# Duncan S Anderson

Duncan S Anderson writes The Prompt Factory newsletter and shares analysis and practical experiments on AI models, agent systems, and the costs and trade-offs of running them. Anderson's posts cover topics such as making agent tools more reliable and evaluating smaller, lower-cost models.

- Profile: https://b2binfluencer.marketing/creators/duncan-s-anderson/
- LinkedIn: https://www.linkedin.com/in/duncansanderson
- LinkedIn followers: 4,004
- Topics: artificial-intelligence
- Based on 13 dated posts, latest 2026-10-08

## Summary

Useful for technical content on agent harness engineering, model costs, and cost-efficient model selection for AI infrastructure buyers.

- Best for: A B2B AI infrastructure company wanting technical content on agent harness engineering and cost-efficient model selection.
- Not for: A consumer AI app growth team seeking paid creator installs through short-form video for a general audience.
- Check first: Confirm terms, disclosure, analytics, and audience data; no rate card or inquiry route is published.

### Takeaways

- Moved a live agent from Claude Sonnet to cheaper Nemotron; it broke immediately.
- Trace review showed failures came from the harness, not just the model.
- A 27B model on a laptop shifts the data-centre debate.

## Main ideas

- **Near-miss model output is an engineering problem, not a reason to buy a bigger model**: Duncan moved a production agent from Claude Sonnet to a much cheaper small model, and it stopped working properly. The obvious first read was that small models cannot sustain multi-step tool use. But traces showed almost every failure was mechanical: how tool calls were named, parameterised and parsed. A harness can absorb those errors, so Duncan built one that expects the model to be "almost right" rather than paying for a cleverer model. A failed model swap can reveal as much about the surrounding code as the model.
- **Cheaper inference is also an energy decision**: Duncan treats token price as a proxy for compute, and compute as a proxy for power, making model routing an environmental choice as well as a budget decision. A fraction of the cost means a similar fraction of the power. The agent experiment shows this choice is available now that a cheaper model can be made to work.
- **A model that runs on a laptop undercuts the case for giant data centres**: Duncan uses an independent benchmark of tool use, planning and multi-step problem solving — the capabilities that matter when building agents rather than chatbots — to make a hardware argument. A 27B parameter open-weight model high on that leaderboard, yet running on an ordinary laptop, challenges the assumption that serious agentic work requires enormous compute and the data-centre and power buildout that assumption justifies. Duncan notes an allegation that Alibaba used Anthropic model outputs to train Qwen, then sets it aside to make the wider claim.
- **Open replication outruns the hard parts of a new idea**: Duncan tried a decision model in a chatbot routing layer, found it faster and cheaper, then watched the idea get copied from several directions within days. The excitement is real, but calibration and tooling remain unsolved. A decision model is useful only if its stated 80% confidence actually means 80%. The spread shows how fast open work moves, not that decision models are production-ready.
- **Analysis over news, ending in one thing to do**: Duncan runs The Prompt Factory as a recurring slot built around a forced view, not a headline feed: a few argued points, links for everything else and one action the reader can complete. Picking three items is harder than picking six because it forces a position — and that position is the value. The format also suits a co-produced or sponsored edition: it already commits to an argument and a test.
- **Model access is becoming a question of nationality**: Duncan follows a model takedown as a policy story rather than a product one, tracing how the order came about and what it means for anyone choosing a provider. The reported sequence — a jailbreak flag from Amazon's CEO, Amazon's stake in Anthropic and the timing just after Dario Amodei argued for tighter government oversight of frontier AI — leads to Duncan's question about access. If getting the best model starts to depend on proving your nationality, that is a structural risk for labs and buyers outside the United States.

## Work to read

- [The cheap model broke. The harness was the problem](https://www.linkedin.com/posts/duncansanderson_claude-sonnet-costs-two-dollars-per-million-activity-7495015318419918848-L4aF)
- [A 27B model on a laptop shifts the data-centre debate](https://www.linkedin.com/posts/duncansanderson_artificial-analysis-has-become-the-closest-activity-7495377736828604416-zoJw)
- [Decision models spread faster than calibration](https://www.linkedin.com/posts/duncansanderson_last-week-i-dropped-a-jev-the-viral-ai-decision-activity-7507749456453296128-x8R5)
- [When a model takedown becomes an AI policy story](https://www.linkedin.com/posts/duncansanderson_washington-switched-off-the-best-model-activity-7471834567147163649-WbFq)
- [A newsletter that picks three things and one action](https://www.linkedin.com/posts/duncansanderson_this-weeks-the-prompt-factory-newsletter-activity-7512430951579373568-7vfN)

## Brand relationships

- **The Prompt Factory** (owned_business)
- **Anthropic** (mention)
- **Nvidia** (mention)
- **Alibaba** (mention)

## Partnerships

No sponsorship formats, media kit, rate card, or official partnership inquiry route appear anywhere in the posts here, so this section documents an absence rather than a menu of options. What the posts here do show is how Anderson publishes: the weekly Sunday newsletter The Prompt Factory, promoted through LinkedIn posts that link out to the full edition. The current newsletter format presents three main stories, one line of links to everything else worth reading, and one suggested action, with the stated aim of being more analysis than news rather than a news digest. For a buyer, the practical consequence is that any commercial arrangement would have to be opened directly through the LinkedIn account listed in the identity section. There is no published specification of placement, sponsorship labelling, pricing, lead time, exclusivity, or review process to plan against, and nothing in these posts shows how a partner would be credited or disclosed. Because these posts contain no sponsored content, affiliate disclosure, or partner credit of any kind, none of this should be read as evidence that Anderson does or does not accept sponsorships. It means only that no offer is documented in these posts. Confirming whether a partnership is possible, and on what terms, requires a source beyond these posts: a media kit, a sponsorship or contact page, or an explicit statement from the creator.

## Alternatives

- Alex Cinovoj: Pick Alex Cinovoj when the collaboration needs a failure-and-repair case study rather than a comparative analysis. The two posts lead with shipped-agent counts and named incidents: a Series B fintech that spent $120K over six months and ended up firing 847 API calls instead of 12, with a $6,200 first-week bill, rebuilt in three weeks at $0.08 per workflow; a Sydney logistics firm at $180K; and a Dutch firm claiming 60% admin reduction. Cinovoj also states the buyer's argument directly: the model is roughly 10% of what makes an agent work, while tools, rules, sandboxes, guardrails, and observability are the rest. For a vendor pitching reliability and observability, that is campaign-ready framing delivered as a story, faster to commission than Duncan S Anderson's trace-by-trace method. Alex is the stronger choice for a provocative narrative post or a worked example drawn from client engagements, and the weaker choice when the deliverable must be a reproducible test a reader can run in an evening.
- Jim Kaskade: Pick Jim Kaskade when the collaboration targets executives rather than working developers. The posts argue that the agent, not the model, becomes the control plane for work, and that whoever owns agent routing eventually owns procurement because invisible lock-in replaces human comparison of alternatives. The other post examines military AI procurement: alignment becoming procurement requirements and audit trails, and the blunt position that the safest military AI is open evaluation and enforceable constraints rather than refusal. That framing suits a vendor selling governance, permissions, and observability to budget holders, and reaches policy and CEO readers that a trace-level engineering post does not. Choose this for a campaign on enterprise orchestration or procurement accountability.
- Mike Miller: Pick Mike Miller when the campaign fits security practitioner coverage better than AI engineering. The roundup collects dated developments, including a CVSS 9.5 Citrix Netscaler issue with tens of thousands of exposed instances, an OpenAI training-run safety process, and a new cyber benchmark index. It gives a sponsor a slot in a recurring news cadence aimed at security teams. The second post moves from vulnerability volume to remediation, showing how vibe-coded applications multiply what teams must triage and walking through an agent remediation flow: understand the issue or dependency, find the safest upgrade, apply the fix, verify it, and open a pull request, with Patch Tuesday recorded at nearly 1,000 vulnerabilities. That post states a partnership with Legit Security and is written as vendor-facing content, establishing precedent for sponsored placement and a format to fit. Use it for a campaign on security remediation, CVE triage, or a named sponsored slot in an existing publication rhythm.

---
AI-assisted summary of public posts. Audience demographics are not available. Engagement appears only where we checked recent posts. Method: https://b2binfluencer.marketing/methodology/