The day my agents said no
On 31 August, two of my AI agents refused an instruction from me, and they were right to.
They were halfway through a batch of twenty documents when I sent a quick correction. It loosened one of my hard rules. To the agents, it looked like a prompt injection, an instruction smuggled into their work. My briefs told them to treat that as an attack. They flagged it, quoted it back to me and finished the job on the original brief.
A good new starter would do the same with a note saying “skip the sign-off this once”. The rules held while I was in a hurry. That is AI management working.
Where my rules come from
I left code on purpose. During my Master of Science at the University of Glasgow, I wrote C# and spent hours hunting for a missing letter. I wanted to spend my time on ideas. I moved to product design and waited for the tools to catch up.
At Healthily, I owned a legacy AI, the Smart Symptom Checker. I added a layer that showed people what the system had heard before they submitted. It became my first rule for any AI. The AI shows what it understood before it acts.
Now AI does the debugging, and I build what I design in Claude Code and Cursor. Vibe coding is the return I planned.
The team I run
Designer Fund’s 2026 survey of more than 900 designers found leaders calling designers “orchestrators”. That is my job every day.
- 91%
of designers use AI in their work at least weekly, up from 54% a year earlier
AI in Design Report, May 2026 - 43%
say their companies now expect working prototypes as deliverables
AI in Design Report, May 2026 - 13
of 115 product design adverts named Claude Code as a tool
My research, September 2026
My team has four levels. I set the goal and sign off every release. One model plans and judges. Faster agents do the scoped work, and plain code guards the edges.
Most of it runs without me. Seventeen scheduled agents work through the day. A run log flags any run that claims success while hiding an error. On its first day, it caught 51 of the 510 runs it recorded.
How I manage my agents
An agent is a strange kind of user. It starts every job knowing nothing and reads everything literally. When a brief leaves a gap, it fills it with something plausible. Nielsen Norman Group now treats agents as users in their own right. So the brief is my interface, and I design it like a form.
I distrust a screen that looks too tidy. On 20 September, a run that had died on a rate limit still showed as succeeded after five seconds. A green run under 30 seconds now counts as a failure.
I design for my own blind spots too. Researchers call it automation complacency. On a busy day, I check less. A change now counts only once it has been read back from the document itself. Every delete becomes a move to a folder I review.
What I learned
Back to 31 August. A worker agent sees only its brief, so a correction halfway through looks like an intruder. I kept the rule and changed my habit. Every rule now goes in the first brief. A correction starts a fresh agent.
That is the job. I decide what the agents may do alone and write it where they cannot miss it. Then I check their work before it ships.
Sources
How we built our multi-agent research systemAnthropic · Engineering
Subagents: each one starts with a fresh, isolated context windowClaude Code · Documentation- Hooks: rules that run before an action and can block itClaude Code · Documentation
- AI in Design Report 2026Designer Fund and Foundation Capital · Survey of 900+ designers
- AI agents as usersNielsen Norman Group · April 2026
- Complacency and bias in human use of automationParasuraman and Manzey, Human Factors · 2010
- AI in Design 2026Designer Fund · May 2026 · The report’s summary, source of the “orchestrators” quote