Simon Willison's 2026 recap has one lesson for automation builders: defining the job is the skill

The most useful sentence in a long AI year recap was about writing a good brief.

What happened

Simon Willison gave the closing keynote at the WeAreDevelopers World Congress North America in San Jose and published annotated slides on September 27. He walks through the year month by month, and a few threads stand out for operators.

• Coding agents crossed a line. He says late 2025 models paired with their agent harnesses went from "often make mistakes" to "reliable enough to use on a day-to-day basis."
• Agents are expensive. He notes it was hard to spend more than $50 on tokens last year, and now "you can actually spend $1,000 in a day doing real work." Companies that pushed maximum token usage later capped it.
• Local models got strong. He calls Qwen 3.8 27B, a 17GB download running on his laptop, "an extraordinary model."
• On what he calls Fable class models such as GPT-6 Astra: if you "clearly define the goal," give "unambiguous instructions about the constraints," and provide the tools, the model will solve the problem "through brute force." His point is that defining goals, constraints and tools "is kind of what software engineering is."

My take

That last point is the whole job in business automation too. When a client project stalls, the model is rarely the blocker. The brief is. "Automate our lead follow up" is not a goal. "Every web lead gets a reply in five minutes, tagged by service, assigned to the right rep, and nothing is sent after 9 PM local time" is.

Three things I take from the recap:

1. Spend the first session writing the definition of done, the constraints and the list of tools the agent may touch. That document is the most valuable artifact of the project.
2. Budget for tokens like any other running cost. Put caps on agent runs before usage surprises you.
3. For sensitive or high volume tasks, local and open models are now worth a real test instead of a default no.

Better models reward teams that can say clearly what they want. That skill is cheap to learn and it compounds.