Start here · hensq
AI agents in real workflows, not just demos.
I test how an agent can move from completing a task to owning a bounded piece of work—then document what succeeds, what fails, and where a person still needs to step in.
All cases are anonymized or synthetic. I do not publish private messages, identities, prices, routes, or unverified business results.
Evidence & field notes
Three useful places to begin
Each card includes an English summary. The detailed field notes are currently written in Chinese.
From completing tasks to owning bounded work
A five-level model for deciding how much responsibility an agent can safely take.
- Responsibility grows from advice to bounded tasks, connected steps, and normal-case ownership.
- Each upgrade needs detectable failure, a stop or recovery path, and a named exception owner.
Measure value beyond model speed
A practical chain from task performance to process, customer, and operating outcomes.
- Measure review, correction, recovery, and waiting time—not model latency alone.
- If evidence stops at workflow performance, do not claim a customer or revenue result.
Treat human handoff as part of the workflow
The triggers and minimum context an exception needs before it reaches a person.
- Stop on missing input, conflicting rules, high-impact action, uncovered cases, or system failure.
- Pass the request, facts, rule version, action taken, reason, owner, and next decision.
Daily build log
The short experiments happen on X.
The site keeps the durable models. X has the day-to-day tests, failures, and corrections.