AI Guides › Skills & Agents
Building An Agent That Reports Instead Of Acting — The Safer First Version
By Nigel Guy · 2 min read
The instinct with a new agent is to build the version that does the whole
job — finds the thing, decides what to do, does it. That's also the version
most likely to go wrong in a way you don't catch until it's already acted.
There's a safer, almost identical version that catches this.
The rule: build the reporting version first — same analysis, same
recommendation, but it tells you what it would do instead of doing it — and
only promote to acting once the reports have earned your trust.
The mechanism
- Build the agent to do everything except the final action. It gathers
information, reasons through it, and produces a clear recommendation:
"I would do X, because Y."
- Run it for real, on real cases, for a genuine stretch of time. Not a
handful of tests — actual ongoing use, long enough to see how it handles
the variety of situations that come up naturally.
- Compare its recommendations to what you'd have actually done. Where
it agrees with your own judgement consistently, that's real evidence.
Where it doesn't, that's exactly the gap worth understanding before
granting it the ability to act.
- Promote narrowly, not all at once. Let it act on the specific case
type it's been most reliably right about, keep reporting-only on the
rest, and expand gradually as the track record grows.
What to skip
Skip jumping straight to an acting agent because the reporting version
"seems obviously right" after a few runs — a few runs is exactly the sample
size that feels convincing and isn't. Skip promoting an agent to act on
everything it reports on just because the promotion for one case type went
well; each case type earns its own trust.
Guardrails
- The reporting phase isn't a delay tactic — it's the actual verification
step. Skipping it to "save time" is skipping the one thing that catches
a bad agent before it costs anything.
- Keep the reporting version running even after some actions are promoted —
it's your ongoing check that the agent's judgement hasn't quietly drifted.
- Anything irreversible stays in reporting-only, regardless of track record,
unless you've deliberately decided the cost of an occasional wrong action
there is genuinely acceptable.
All 751 AI guides · JulieMango plans from £17/mo