Josue Hernandez creates practical guides and tools for using AI in sales and revenue operations, including agent workflows, model-cost comparisons, and lead-generation systems.
Useful for practical AI sales and revenue operations guides, agent workflows, cost comparisons, and lead-generation systems.
Best for
A B2B sales-automation brand hosting a RevOps workshop on secure, cost-effective AI agents.
Not for
A consumer wellness or DTC brand seeking paid lifestyle posts with proven engagement and audience demographics.
Key takeaways
Separate bots on one account are not a security boundary; they share a cloud computer.
An approval predicate returning None can be read as approval.
Models priced by workflow; crossover near 300,000 cached tokens with light output.
Check first: Confirm sponsorship format, rates, and documented audience data covering employer and seniority.
Main ideas
Judge AI models by workflow cost, not leaderboard rank · Read the vendor's limits page — it may contradict the tutorials · Build the safety layer before the interesting part
Josue Hernandez creates practical guides and tools for applying AI to sales and revenue operations.
Their work focuses on how agents handle recurring sales tasks, what they can access, when people need to approve actions, and what repeated model use costs.
The posts also show work on lead-generation systems, including Dapta Forms and a workflow for turning YouTube videos into LinkedIn infographics.
Their examples make them worth following if you want concrete designs and calculations rather than broad claims about what AI can do.
Hernandez connects AI features to the operational details that determine whether they’re useful: permissions, failure handling, approval gates, workload limits, and cost.
In one post, they cite vendor documentation warning that separate bots don’t create a security boundary, then explain the shared computer and browser sessions behind that warning.
In another, they lay out five revenue routines and the platform limits that constrain them.
Their model comparisons show the arithmetic and, in the benchmark post, disclose that the first two runs weren’t blind and therefore have no correctness score.
This mix of concrete implementation detail and stated limits is the clearest reason to read their work.
Ask whether they have signature books, shows, or earlier work that explains name searches; the posts here don’t establish that.
Josue Hernandez's recurring argument is that the practical questions about AI in sales concern cost, safety and scheduling, not which model scores highest.
Read more
Across eight posts, they work from vendor documentation and arithmetic rather than benchmarks or tutorials, design agent systems so nothing sends or writes without a human gate, and treat a scheduled, self-contained routine — not a chat window — as the thing worth building. Every guide ships with the calculator, fixtures or settings behind its numbers, and every post states where the claims stop.
Judge AI models by workflow cost, not leaderboard rank
Josue Hernandez argues that the usual head-to-head comparisons answer the wrong question: the price sheet hides the number that actually drives the bill.
Read more
Input and output rates can be identical across two models while cache reads — the cost of rereading the same files on every loop — differ enormously.
They recommend working out the arithmetic for the specific workflow: the size of the cached corpus, the output load and the number of loop repetitions.
From that, they distil a rule they say you can hold in your head: the expensive model reads and decides; a cheaper one does the typing.
One post separates a model's input price from its cache-read price: reads cost $0.25 per million tokens — about 2.5% of its input rate, versus roughly 10% for other Claude models.
It then prices two workflows.
Loading a 745,000-token archive and reasoning across it costs $21 on Claude Fable 5.1, $22 on Opus 5 and $45 on the earlier version; writing the memos in the same session reverses the result, at $9.76 on Fable 5.1 versus $8.26 on Opus 5.
The crossover comes around 300,000 cached tokens with light output.
A companion post makes the same point about framing: both models list at $10 in and $50 out per million tokens, so a single request ties to the cent, while an agent loop over the same pipeline diverges by 7% to 46%, depending on token mix.
Read the vendor's limits page — it may contradict the tutorials
Several posts closely read a single sentence in product documentation that undercuts what most guides teach.
Read more
The method is consistent: find the page the tutorials skipped, quote it, then spell out the operational consequences for a team that followed those tutorials. Josue Hernandez treats vendor documentation as primary evidence and presents the correction as something the vendor states plainly, rather than as a flaw they discovered.
One post describes running the recommended setup — one Chief of Staff bot with five specialists for inbox, calendar, CRM, meetings and reporting — then opening the security page and finding the line "Do not use separate Bots as a security boundary."
The reason: every bot on an account shares one cloud computer, one filesystem and one set of browser sessions.
Signing one bot into a CRM gives every other bot that session, including bots created later.
The stated fix is to connect a limited role rather than an owner login.
A second post reports a documentation line saying a task will stop mid-run with no resume and no undo.
It also flags a flaw in which the Agents SDK wraps approval decisions in bool(): a predicate returning None becomes False, which reads as no approval needed and lets a guarded tool fire without asking anyone.
Build the safety layer before the interesting part
The design position behind the documentation readings: an unattended agent needs stop handling and approval gates from the outset, not as an add-on.
Read more
Josue Hernandez describes delaying feature work to build a handler for being stopped and a gate that fails closed — so when something goes wrong or is ambiguous, the default is to refuse, not proceed. This matters because revenue agents run overnight, when nobody is watching; a silent failure can look like success until money or data has already moved.
After finding the mid-run stop behaviour, Josue Hernandez says they stopped building the interesting part and built the boring part first: a stop handler and a gate that fails closed, wrapped around one operator agent with five weekly revenue jobs.
Other pieces follow the same approach: inbound triage that returns a CRM-ready record at six cents a lead, an overnight pipeline review, and SPF and DMARC checks resolved from inside a sandbox before anything sends.
The same post notes four issues filed about the boolean approval behaviour in a single week.
The routine, not the chat window, is the unit of agent value
Josue Hernandez argues that most people still use a cloud-hosted agent as a messaging window, when value comes from a routine: one job, owned by one bot, triggered on a schedule or by an event, and running whether or not a person is present.
Read more
Their specifications cover the trigger, steps, approval rule, output, connectors and limits, and account for the platform's published ceilings before production.
One post specifies five revenue routines: nightly account research that stops at a review list and never sends; inbound lead triage that scores, routes and drafts, with every write gated; a Monday paid media report showing current value, proposed value and expected impact; a Friday account health watch list with thresholds that do not drift; and an event-triggered deal desk pack that assembles and stops before Slack. The post also lists the constraints: 50 routines per bot, 50 bots per account, only 20 run records per routine, one computer-use task per bot at a time, so five browser sources run in series; $20 a month solo and $40 per user for a team; and a trial that is a usage credit, not seven days — one long job can exhaust it.
Publish the numbers and files, and state where the claims stop
The guides ship the calculator, prompts, fixtures and per-routine settings alongside the conclusion so someone else can rerun the work.
Read more
Josue Hernandez also names the limits of their own tests in the post itself — a non-blind run, a model they had no key for, a comparison with no correctness score. Readers can check the claims and see what was tested before acting on them.
The model-cost post includes a 40-line cost calculator with every assumption exposed, an effort setting per routine and a note that low effort breaks research.
The sales-task benchmark logs every tool call, turn and cached token; ships the fixtures, keys and five prompts so the bench can be rerun clean; and states that the first two runs were not blind, so it includes no correctness score.
A later agent post says no API key was available and nothing was run live, and puts that information on page one.
Strong ads pointed at a weak form still collect junk leads
The marketing-side argument pairs acquisition with capture: paid media playbooks and a qualifying form come together because traffic sent to a form that does not screen or route produces volume rather than pipeline.
Read more
The numbers are the author's published campaign figures, not independently audited results.
A launch post for Dapta Forms describes it as free, open source, MIT licensed and self-hostable: a form that qualifies, scores and books a meeting on its own.
It offers 15 paid media playbooks to the next 30 signups.
Listed contents include a cost per lead reduced from $50 to $18, webinar signups at $2 each, a five-minute window said to decide whether a lead ever answers, eight hook templates, how Meta auction and learning phase actually behave, and a triage protocol for a campaign that dies overnight.
The post argues for shipping both halves at once because a phone field alone costs 32% of signups.
The delivery format shifted from keyword-gated promotion to technical build-alongs
The earliest post in the set addresses a different audience and uses a different register from everything that follows: content creators struggling with consistency, a video-to-infographic pipeline, and access gated behind commenting a keyword and reposting for priority access.
Read more
Later posts target operators building sales agents, lead with a contradicting documentation line or a piece of arithmetic, and close with a free downloadable guide and working files. The shift reads as a change in focus rather than a settled second specialism.
The early post sketches a four-stage system that automatically pulls a YouTube transcript, extracts key points and data, generates a professional-grade infographic with Google's Nano Banana Pro, then writes a caption and schedules the LinkedIn post.
It closes with a comment of "INFOGRAPHIC" and a repost for priority access.
Later technical posts instead list shipped assets, including a Python skeleton with a stop handler and a gate that fails closed, a read-only analyst sub-agent, a hook that blocks CRM writes until approved, and a connect-now / gate-it / never-connect table.
Answers the question most agent-building tutorials skip: if I split my automations into separate bots, is that actually isolation?
Read more
A short case study in reading the vendor's own limits page instead of the tutorial.
Josue says they were already running the setup most guides recommend — one Chief of Staff bot and five specialists covering inbox, calendar, CRM, meetings and reporting — when they opened xAI's security page and found the line "Do not use separate Bots as a security boundary."
The explanation is concrete: every bot on an account shares one cloud computer, one filesystem and one set of browser sessions, so signing a single bot into a CRM exposes that session to every other bot, including ones created later.
The post contrasts this with tutorials that tell people to save logins into each bot and call it secure, then outlines the guide's contents, including a connect-now / gate-it / never-connect table and the fix of connecting a limited role rather than an owner login.
For brands planning a campaign
Useful teaching material for a workshop on agent security design: the vendor documents the failure mode rather than an attacker discovering it. It also sets expectations for a brief: Josue quotes primary documents and provides detailed sourcing.
A clear example of an agent failure that only shows up in unattended runs — the kind of detail a newcomer might not think to ask about.
Read more
Read it to see what fail-closed design means in code.
Josue reports finding a single line in OpenAI's documentation stating that in the API the task will stop, with no resume and no undo, and that whatever the agent already sent has already been sent.
A second problem follows: the Agents SDK wraps approval decisions in bool(), so a predicate returning None becomes False, which reads as no approval needed, meaning a guarded tool fires without anyone being asked; four issues were filed on that in a week.
The response was to build the safety layer first — a stop handler and a gate that fails closed — around one operator agent with five weekly revenue jobs, an inbound triage routine returning a CRM-ready record at six cents a lead, an overnight pipeline review, and SPF/DMARC checks run from inside a sandbox before anything sends.
Josue says they had no API key and ran nothing live; ask for a live test before deployment.
For brands planning a campaign
Strong workshop material on failure modes and approval design. Ask whether the architecture has been tested in a live deployment; the post reasons from published documentation.
Answers "what does an AI agent actually do all day in a sales team, and what stops it?"
Read more
The most concrete routine-by-routine specification in the material.
The post argues that the routine is the unit of value: one job, owned by one bot, triggered on a schedule or an event, running whether or not a person is present.
Josue specifies five for revenue work: nightly account research that stops at a review list rather than a send, gated inbound lead triage that scores, routes and drafts, a Monday paid media brief showing current value, proposed value and expected impact, a Friday account health watch list with fixed thresholds, and an event-triggered deal desk pack that assembles and stops before Slack.
The design numbers are listed too: 50 routines per bot and 50 bots per account, only 20 run records per routine, one computer-use task per bot at a time (so five browser sources run in series), $20 a month solo and $40 per user for a team, and a free trial that is a usage credit rather than seven days, which one long job can exhaust.
For brands planning a campaign
Directly usable for a build-versus-buy discussion: it prices the setup and names the ceiling that bites first. The routine list can also serve as a workshop worksheet for mapping sales tasks to agents.
Explains why two models with identical sticker prices produce very different bills, with the arithmetic shown.
Read more
Start here to learn how to reason about AI costs rather than which model is "better."
The central distinction is between input and output rates, which Josue says are unchanged at $10 and $50 per million tokens, and cache reads, which are $0.25 — about 2.5% of input, against roughly 10% for other Claude models.
Instead of offering a general verdict, the post works through two workflows: loading a 745K-token archive and reasoning across it costs $21 on Fable 5.1, $22 on Opus 5 and $45 on the earlier version, while writing memos in the same session costs $9.76 on Fable and $8.26 on Opus.
The crossover is around 300K cached tokens with light output, above which the cheaper cache reads win.
Josue's rule: the expensive model reads and decides while a cheaper one does the typing, implemented as a line of frontmatter per sub-agent in Claude Code.
The accompanying guide includes a 40-line cost calculator with its assumptions exposed, per-routine effort settings, a read-only analyst, and a hook that blocks CRM writes until approved.
For brands planning a campaign
Useful for a workshop module on costing agent workloads before deployment, and as a sample of recommendations built around a stated crossover point rather than a blanket preference.
Shows what a sales-relevant model comparison looks like when it is built from real tasks and honest about its limits.
Read more
Read it for the test design, not just the conclusion.
GPT-6 Astra and Claude Fable 5.1 are run through the same five sales tasks on identical files, prompts and scorer: triaging 60 inbound leads into strict JSON with ten planted edge cases; a deal risk memo from 25 call transcripts at 167K tokens with no chunking; deduplicating a 300-row CRM over MCP without merging the wrong people; a full outbound kit including SPF, DKIM and DMARC verdicts on three domains; and scraping a paginated directory containing three scraper traps.
Josue notes that head-to-heads published the previous week reported zero token counts, which is what buyers actually pay for, so this benchmark logs every tool call, turn and cached token.
On a single request, the two bills are identical to the cent; on an agent loop, the gap runs from 7% to 46% depending on token mix.
The post says the first two runs were not blind, so it includes no correctness score, and that fixtures, keys and prompts ship so others can rerun it clean.
For brands planning a campaign
A useful reference for briefs that require evidence-quality detail: the methodology limits are disclosed alongside the results. Ask for a blind correctness score before using the benchmark to support claims about accuracy.
The clearest look at Josue's own product work and demand-generation side, rather than analysis of other vendors' tools.
Read more
Read it to judge fit for campaigns about lead capture and qualification.
The post launches Dapta Forms, described as free, open source, MIT licensed and self-hostable: a form that qualifies, scores and books a meeting on its own.
The promotion offers the next 30 signups 15 paid-media playbooks, framed as the same files the team opens on a Monday morning, and lists the contents: a cost per lead reduction from $50 to $18, webinar signups at $2 each, a five-minute window said to decide whether a lead answers, eight hook templates, an explanation of Meta auction and learning-phase behaviour, and a triage protocol for a campaign dying overnight.
The argument tying the two halves together is that strong ads pointed at a weak form collect junk leads, with the claim that a phone field alone costs 32% of signups.
For brands planning a campaign
Relevant to campaigns touching lead capture or form design, and a launch pattern that pairs a free tool with a gated playbook. The performance figures are the author's own; ask for independent verification before using them in campaign claims.
Shows what Josue does with generative image tools, complementing the model-cost and agent posts.
Read more
It is the clearest content-marketing example in the material.
The post sketches a four-stage system: choose a YouTube video and pull its transcript automatically at any length or topic; process the content to extract key points and data and structure them for visual impact; generate a professional-grade infographic with Nano Banana Pro and automatic layout; then write a caption and schedule the post to LinkedIn.
Josue connects it to a common problem — creators struggling with consistency while transcribing, designing and formatting — and frames the system as a response to that workload.
Access is gated behind connecting and commenting "INFOGRAPHIC," with a repost for priority access.
For brands planning a campaign
Useful for a campaign involving repurposing long-form video into social assets, and as a sample of Josue's lead-magnet mechanics. A brand would need to decide whether to adopt the comment-and-repost gate as a distribution tactic.
Instructional format: build-along guides with shipped files7
Choosing AI models by workflow cost, not benchmark rank3
Reading vendor documentation for security and failure modes2
Scheduled revenue routines as the unit of agent value1
Lead generation, forms and paid media operations1
Generative-media content automation1
Choosing AI models by workflow cost, not benchmark rank
Three posts argue that the usual “which model is better” debate relies on the wrong evidence.
Read more
One compares cache-read pricing — $0.25 per million tokens on Claude Fable 5.1, described as 2.5% of its input rate against 10% for other Claude models — and calculates the cost of reading a 745,000-token archive: $21 on Fable 5.1, $22 on Opus 5 and $45 on the previous version.
Writing the memos afterwards in the same session flips the result: $9.76 on Fable 5.1 against $8.26 on Opus 5.
Josue Hernandez puts the crossover at roughly 300,000 cached tokens with light output and reduces the rule to “the expensive model reads and decides, a cheaper one does the typing.”
A second post benchmarks GPT-6 Astra and Fable 5.1 on five sales tasks — triaging 60 inbound leads into strict JSON, writing a deal-risk memo from 25 call transcripts at 167,000 tokens, CRM deduplication over MCP, building a full outbound kit with SPF, DKIM and DMARC verdicts, and scraping a paginated directory — logging every tool call, turn and cached token.
Because both list at $10 input and $50 output per million, single requests tie to the cent, while agent loops diverge by 7% to 46%, depending on token mix.
A third post makes the same argument about framing: the winner is not the smarter model but the one that stays cheap when it reads the same pipeline every day for a month.
Publishes a 40-line cost calculator with every assumption exposed, ships test fixtures and prompts so the comparison can be rerun, and describes setting model effort levels by routine.
They disclose that their first two comparison runs were not blind and include no correctness score.
For brands planning a campaign
For a workshop on cost-effective sales agents, this is the strongest entry point. Hernandez starts with the buyer’s bill, not a leaderboard, and walks sales-ops teams through the crossover arithmetic for their own token patterns.
Reading vendor documentation for security and failure modes
Two posts closely read documentation that undercuts common practice.
Read more
One reports that Grok Bot’s security page says not to use separate bots as a security boundary: every bot on an account shares one cloud computer, filesystem and set of browser sessions.
Signing one bot into a CRM gives every current and future bot that session, including bots created next month.
They contrast this with tutorials that teach saving logins into each bot and describe their own fix: connect a limited role, never the owner login.
The other reports that GPT-6 Astra can stop an agent mid-run with no resume and no undo, and that the Agents SDK wraps approval decisions in bool(), so a predicate returning None becomes False and is read as no approval needed, letting a guarded tool fire unasked.
They say four issues were filed about this in a single week, and that they built the stop handler and a gate that fails closed before building the interesting part.
They disclose that they had no API key for the model in the second post, that they did not run anything live, and that this is stated on page one.
For brands planning a campaign
This is diligence material sales-automation buyers rarely get from a creator: named documentation behaviour, its operational consequence and a specific mitigation. It supports a workshop module on where agent approvals and credentials break.
Scheduled revenue routines as the unit of agent value
One post argues that the valuable unit of a cloud-hosted agent is the routine: one job, owned by one bot, triggered on a schedule or after an event, whether or not the operator is present.
Read more
It specifies five routines for revenue work: nightly account research that stops at a review list and never sends; inbound lead triage that scores, routes and drafts, with every write gated; a Monday paid media brief showing current value, proposed value and expected impact; a Friday account-health watch list with thresholds that do not drift; and an event-triggered deal desk pack that assembles and stops before Slack.
Each is specified end to end — trigger, steps, approval rule, output, connectors and limits.
The design constraints are quantified: 50 routines per bot, 50 bots per account, 20 run records per routine, one computer-use task per bot at a time, $20 a month to run it solo and $40 per user for a team, with a free trial that is a usage credit rather than seven days.
For brands planning a campaign
Teams planning agent operations can lift the routine template — trigger, approval rule, output, connector, limit — and use the retention and concurrency ceilings to size an implementation before committing budget.
One post speaks as part of the Dapta team at the launch of Dapta Forms, an open-source, MIT-licensed, self-hostable form that qualifies, scores and books meetings on its own.
Read more
It offers 15 paid media playbooks to the next 30 signups, covering how to reduce cost per lead from $50 to $18, get webinar signups at $2 each, use a five-minute window to decide whether a lead ever answers, apply eight hook templates, understand how the Meta auction and learning phase behave, and run the triage protocol when a campaign dies overnight. The case for pairing the playbooks with the tool: ads pointed at a weak form collect junk leads, and a phone field alone costs 32% of signups.
For brands planning a campaign
Useful for campaigns covering top-of-funnel mechanics rather than agents: they can speak to form design as a conversion lever and cite specific paid-media numbers. These are their own published figures, not audited results.
Instructional format: build-along guides with shipped files
Every post reviewed follows the same delivery pattern: an opening finding traced to a quoted document or arithmetic, followed by a numbered list of what the accompanying guide contains — a Python skeleton with a stop handler and a gate that fails closed, inbound triage returning a record a CRM accepts at six cents a lead, a 40-line cost calculator, a read-only analyst sub-agent, a hook that blocks CRM writes until approved, or a connect-now / gate-it / never-connect table.
Read more
The closing call to action usually offers a free guide download; the earliest post instead asks readers to comment a keyword and repost for priority access.
Leads with a counterintuitive or contradicting finding, shows the arithmetic or quoted documentation line, then lists the shipped assets and gates access behind a single call to action.
For brands planning a campaign
This is a clear collaboration opportunity.
The deliverable is a written guide with working files rather than a sponsored post, making a joint build-along the natural co-marketing format.
The asset lists also make the scope easy to price.
The earliest post in the set covers a different subject and audience.
Read more
It describes a system that turns any YouTube video into a professional-grade infographic using Google’s Nano Banana Pro: automatically pulling the transcript, extracting key points and data, generating the infographic, then writing captions and scheduling the post to LinkedIn. It frames the problem as creator consistency — hours spent transcribing, designing and formatting — and closes with a comment-keyword and repost mechanic rather than a downloadable guide.
This post appears earliest in the account’s sequence; the difference in register and depth may reflect a shift in focus rather than a second specialism.
For brands planning a campaign
Shows range into creative production and social scheduling, useful for campaigns that touch content marketing as well as sales operations. It is the least technical and least specific post in the set, so ask whether it represents their current work before using it as a brief.
Josue Hernandez’s posts describe one business affiliation: Dapta.
Read more
The creator describes the Dapta Forms launch in the first person and links to the team’s paid-media system, but doesn’t state an ownership or employment role; ask about their role. Other relationships are product mentions: AI models from Anthropic, OpenAI, xAI, and Google, which are the subjects of the creator’s cost, security, and agent-reliability analyses.
Dapta
Disclosed partnership
A post announces the launch of Dapta Forms and gives away 15 paid-media playbooks described as the files “our team opens on a Monday morning,” placing the creator inside the Dapta team.
Read more
The post doesn’t say whether the creator owns the business or works for it; ask about their role. It also notes that the form tool is open source and MIT licensed, and points to daptaforms.ai.
Original disclosure
We just launched Dapta Forms, and to celebrate, we’re giving away the exact system we use to run our paid media at Dapta.
Claude Fable 5.1, Opus 5, and Claude Code are analysis subjects, not partners.
Read more
The creator compares per-token input, output, and cache-read pricing across the models, calculates a cost crossover at roughly 300K cached tokens, and recommends using expensive models to read and decide, while a cheaper model handles drafting — in Claude Code, one line of frontmatter per sub-agent.
GPT-6 Astra and the OpenAI Agents SDK are analysis subjects.
Read more
The creator reports finding a documentation line stating that a task stops with no resume or undo, and an SDK approval wrapper where a predicate returning None is read as no approval needed, so a guarded tool fires unasked.
Both shape a build that fails closed.
The creator says they had no model key and ran nothing live.
The creator quotes xAI’s security page, which says separate bots aren’t a security boundary because every bot on an account shares one cloud computer, filesystem, and browser sessions. They list the design limits that shape routine work: 50 routines per bot, 50 bots per account, 20 run records per routine, one computer-use task per bot at a time, and a monthly price of $20 solo or $40 per user for a team.
Nano Banana Pro is the image model in a YouTube-to-infographic system the creator says they built: it pulls a video transcript, extracts key points, generates the infographic, then writes and schedules a LinkedIn post.
Read more
The model is a tool in the system, not a stated partner.
Josue Hernandez writes about running AI agents for sales work.
Read more
Read the material in three passes: start with the model-cost posts, which argue that the bill, not the leaderboard, determines which model to use; move to posts quoting vendor documentation to show where agent approvals and credentials break; then read the routine specifications that turn those lessons into scheduled jobs with approval gates.
Each piece offers one finding, the arithmetic or quoted line behind it, and a list of accompanying files, usually in a free guide.
Read the early video-to-infographic post last: it targets content creators and marks a different phase of the account.
Alex Cinovoj, Cole Medin and Mike Futia are useful follow-ons for production-agent failure stories, autonomous software-factory building, and AI creative production for DTC brands, respectively.
Alex Cinovoj also writes about the boundary between AI demos and systems that survive in production, drawing on shipped agents rather than benchmarks—the same ground Josue Hernandez covers from the vendor-documentation and cost side.
Read more
Both treat the unglamorous parts of an agent as decisive.
Alex Cinovoj argues from enterprise outcomes, citing six-figure retainers that produced no deployment and a fintech rebuild that cut a workflow's cost to eight cents.
The proposed remedy is a role: the production AI operator.
Josue Hernandez stays closer to the mechanism, quoting a specific documentation line or cost crossover and shipping the calculator, fixtures or settings behind the claim.
Cole Medin builds autonomous agent systems in public and teaches through working artifacts, much like Josue Hernandez's guides, which include calculators, prompt files and per-routine settings.
Read more
Both argue that reliability comes from the harness around the model, not the model itself.
Cole Medin focuses on autonomous software development: a GitHub issue in, a pull request out, a manager agent overseeing parallel laps, and an open-source factory designed to work out of the box. Josue Hernandez applies similar ideas to sales and revenue work, adding a cost and permissions lens: what a loop bills, what every bot on an account can reach, and where a run stops.
Mike Futia publishes AI playbooks for marketers built around a single tool and a set of named protocols, and gates the full guide behind a keyword comment—the same delivery pattern as the earliest post in Josue Hernandez's set.
Read more
Both target people seeking a repeatable operational routine for a new model.
Mike Futia's posts are visual: turning one product photo into campaign-ready images, swapping models for persona tests, and relighting fatigued ads. Josue Hernandez focuses on agents, integrations and token costs, showing the arithmetic behind the work.
Josue Hernandez posts on LinkedIn to about 8,000 followers.
Read more
A sample of eight posts shows the subject matter buyers would be investing in: model-cost arithmetic for AI agents, approval gates and failure modes in agent design, scheduled revenue routines, a five-task sales benchmark across two models, and a launch campaign for a form product.
The topics point to an audience interested in building or buying AI tools for sales and revenue operations, but the sample does not establish who follows or reads the posts.
The posts do not report impressions, engagement rates, follower growth, geography, company size, job function, or seniority.
There is no audience survey or first-party list, so follower count and post text cannot confirm that readers hold RevOps titles, work at mid-market or enterprise companies, or have budget authority.
For a campaign targeting sales-automation buyers, ask for documented audience data covering employer, occupation, and seniority.
Hernandez offers a channel for practitioner-level reach on AI agent systems for revenue work, with a modest LinkedIn following.
Claims about reaching decision-makers remain unverified.
Request recent post-level results for topics closest to the campaign—the agent-guardrail and cost-comparison posts—including impressions, engagement rate, and follower change over the same window.
If the campaign requires a verifiable share of revenue-operations leaders, ask for documented audience data before treating this as a fit.
LinkedIn followers
8,052
Follower count shown on the creator's LinkedIn profile page · 2026-10-09
Josue Hernandez's posts show no paid sponsorships or brand-partnership formats.
Read more
They promote free guides and tools, typically asking readers to comment a keyword, connect, or repost to receive the resource.
One post announces Dapta Forms, a free, open-source, self-hostable form product that qualifies, scores, and books meetings on its own, and offers the first 30 signups 15 paid-media playbooks.
For a brand considering a campaign, ask directly how partnership inquiries are handled, what placements cost, which deliverables are available, and how sponsored content would be labelled.
A B2B sales-automation company seeking a workshop on secure, cost-effective AI agents for sales pipelines is a good fit, provided the audience consists of sales-operations, RevOps, or automation professionals who will build or supervise agents—not general executives.
Read more
Josue Hernandez’s posts are unusually relevant.
They parse vendor documentation for implementation-critical details: separate bots on one account are not a security boundary because they share a cloud computer, filesystem, and browser sessions; an approval predicate returning None can be read as approval.
They price models by workflow rather than leaderboard, separating headline rates from cache-read costs and placing the crossover near 300,000 cached tokens with light output.
They specify five revenue routines—nightly account research that stops at a review list, gated inbound lead triage, a Monday paid media brief, a Friday account-health watch list, and an event-triggered deal desk pack—and give the platform ceilings and monthly costs that shape the design.
That combination suits a working session for a sales-automation vendor: cost arithmetic, failure modes, approval gates, and routine specifications.
The fit is weaker for a broad demand-generation campaign.
The LinkedIn following is about 8,000, but impressions, engagement, demographics, and job-title data are unavailable.
There is also no published sponsorship format or rate card.
Price the engagement as a practitioner workshop with source-backed teaching, and ask for post-level performance data on the agent and cost posts before adding promotion.
When to choose a specialist
For a consumer wellness or DTC product brand seeking paid lifestyle posts across LinkedIn and Instagram, with proven engagement and audience demographics, this is a poor fit.
Read more
Josue Hernandez’s posts are for people evaluating AI tooling for revenue work: model cache-read arithmetic, agent security boundaries, fail-closed gates, CRM writes, SPF/DKIM/DMARC checks, paid-media cost per lead, and routine design.
One earlier post covers a video-to-infographic workflow, but the rest focus on technical B2B sales operations, with no consumer lifestyle angle.
The audience details list only a LinkedIn follower count, about 8,052; impressions, engagement rates, geography, company size, job function, seniority, and survey data are unavailable.
The posts also publish no sponsorship offer, rate card, media kit, or inquiry route.
They promote Hernandez’s own free guides or tools, with access gated through comment, connect, or repost mechanics.
Ask whether sponsorships are available and request audience data and commercial terms before considering a consumer campaign.
A worked campaign example
Campaign angle for a workshop for sales-operations or RevOps builders: a 90-minute working workshop, “Cost and Containment: Designing Sales Agents That Fail Closed,” for 20–30 sales-operations or RevOps builders, plus a takeaway kit.
Positioning: most agent projects start with the interesting workflow; this session starts with the bill and the failure.
Opening: 10 minutes on why a per-request price comparison misleads.
Walk through the cache-read arithmetic—same input and output rates, very different bills once an agent rereads the same pipeline—and the stated crossover near 300,000 cached tokens with light output.
Have each team mark its recurring workflow as read-heavy or output-heavy and choose which model should read and decide versus which should do the typing.
Core module one: security boundaries.
Show the sentence in vendor documentation stating that separate bots are not a security boundary, explain shared filesystem and browser sessions, and give the fixing pattern: connect a limited role, never an owner login; use the connect-now, gate-it, never-connect table for every system touched.
Core module two: failure modes.
Work through a mid-run stop with no resume and no undo, and an approval predicate that returns None and is read as approval; have teams identify where their current design fails open.
Core module three: routine design.
Use the five revenue routines as templates—nightly account research, lead triage, media brief, health watch, and deal desk pack—and require every routine to name its trigger, steps, approval rule, output, connectors, and first limit.
Include the ceilings: 50 routines per bot, 50 bots per account, 20 run records per routine, one computer-use task per bot at a time, and the solo and team monthly costs.
Close with a worked exercise: take one live sales workflow, produce a one-page routine spec, and name the stop handler and gate that would keep it from acting unsupervised.
Takeaway kit: cost calculator, fail-closed checklist, routine worksheet, and source list for the vendor documentation quotations.
Success measures: every team can price one recurring workflow, name one failure mode and its control, and leave with a gated routine spec ready for staging.
Next step: pilot the lowest-risk routine on test data, with CRM writes blocked until approved, before any production connection.
Ask for the published files and a short pre-workshop call to confirm which examples can be reproduced in the buyer’s environment.
Negotiate the workshop directly and agree on disclosure and data-use terms separately.
Choose Alex Cinovoj when the workshop is about making agents survive production, not about designing the cheapest way to run them.
Read more
The posts build on failure stories and rollout numbers: an enterprise "AI transformation" that spent $120,000 over six months, produced 847 API calls where 12 were expected, and ran up a $6,200 bill in week one before being rebuilt in three weeks; a taxonomy of four ways agent projects stall (large consultancies selling decks, fast prompters who cannot debug a 3am break, researchers who fine-tune what a prompt would solve, and in-house developers who treat probabilistic systems as deterministic software); and small-business rollouts with 90-day payback, 60% admin reduction, and a euro figure of benefit in year one. That gives a session aimed at executives deciding whether to rescue a stalled agent program—or a team that needs help scoping a workflow, wiring integrations, instrumenting observability, handling edge cases, and owning ROI—a stronger spine.
The tradeoff is granularity and sourcing.
Cinovoj argues at the level of a role—the production AI operator—and offers illustrative cases rather than a method a team can rerun.
There is no model-by-model cache-read arithmetic, crossover point for when one model becomes cheaper than another, or per-routine specification template.
A sales-operations team leaving the session would have the diagnosis and mandate, but no cost calculator or gate design.
The posts also use self-promotional framing and make unverified performance claims, so buyers should ask for the underlying engagements before treating the figures as evidence.
Cole Medin is the better choice when the agents under discussion build software rather than run sales operations, or when the session needs a reliability method the buyer can install.
Read more
The posts describe an "AI software factory": a GitHub issue goes in, a pull request comes out, each lap builds one issue, tests the app as a user would, runs an extensive code review, and merges only on green CI, with parallel laps and a manager agent that checks in every couple of hours and pings a human when something is genuinely stuck.
The reusable lesson is how to buy reliability: plan one release of 10 to 20 tickets up front, state what is out of scope, and keep holdout test scenarios hidden from the builder, because an agent that can see the tests will engineer the code to pass them.
Medin also documents converting that year-long experiment into an open source factory alongside Archon, giving a workshop something to download rather than only a principle to follow.
The tradeoff is domain distance.
Medin's loop is about code correctness and test integrity; Hernandez's is about billing, security boundaries, approval gates, and named revenue routines.
A sales-automation team would still need to get the cache-read cost comparison, the warning that separate bots are not a security boundary, and the routine specification template (trigger, steps, approval rule, output, connectors, first limit hit) elsewhere.
Medin's most specific material is also tied to one ongoing build, so buyers should ask how much of it reproduces outside that stack.
Mike Futia is the better choice only if the buyer's decision shifts from agent design to visual creative production.
Read more
The two posts cover turning one product photo into 50-plus campaign-ready images with Nano Banana Pro and Nano Banana 2: virtual photoshoots, instant persona swaps without booking talent, pixel-perfect text rendered inside the image in any language, relighting winning ads to fight fatigue, and localization, packaged as a seven-protocol playbook for DTC brands and agencies replacing $10,000 shoots that die in three days.
For a workshop on secure, cost-effective AI agents for a B2B sales pipeline, Futia covers none of the decision.
The posts contain no security boundary, approval gate, CRM write control, model cost per workflow, or sales routine.
Choosing Futia means shifting the topic from revenue operations to creative operations; the two are not interchangeable for the same campaign.
Josue Hernandez has eight LinkedIn posts observed on October 9, 2026.
The posts cover model price and cache-read comparisons, agent security and approval failures found in vendor documentation, scheduled revenue routines, a lead-capture form launch at Dapta, and a video-to-infographic workflow.
Individual post dates do not appear in the posts, so the order in which the work appeared is approximate.
The posts here do not include videos, transcripts, code repositories, a media kit, a rate card, or audience data beyond a LinkedIn follower count.
The cost, benchmark, and campaign figures quoted are the author's own and were not independently verified, and the vendor documentation appears as quoted in the posts rather than as documentation read directly.
The posts do not identify a signature book, show, or earlier career work.