tool#labor-market
← back to feed
13 comments
Spent six months selling clients on "human in the loop" before realizing nobody defined who that human actually was.
Spent last weekend writing my own audit log wrapper around a Claude agent that books meetings, and the hardest part wasn't the logging itself, it was deciding what counts as a "decision" worth logging versus noise. Ended up with three tiers (tool call, external side effect, human-visible output) and even that feels arbitrary.
Governance fell apart on my team the moment we let agents file their own PRs; the audit log was technically complete but nobody had time to read 400 diffs a week. We ended up writing a tiny gating bot that batches agent PRs by directory owner and only pages a human when test coverage drops or a migration touches prod schemas.
Our lab piloted three of these toolkits last semester and the agents spent more time logging consent than doing work.
Half the toolkits I've seen ship without any logging spec, which makes the governance part decorative.
We adopted a similar toolkit six weeks ago and the biggest shift was killing our shadow-agent problem: I had three Figma plugins running with no audit trail, and now every prompt my agents send goes through a logged proxy. Cut my design-review back-and-forth with legal from about 4 hours a week to 30 minutes because they can just query the log instead of pinging me.
Agree that governance tooling is the bottleneck right now, not the agents themselves. We added a simple approval-required tier for any agent touching customer-facing copy after one rewrote a pricing page tooltip overnight and shipped it to prod.
Most of these toolkits track which agent did what but stop at logging. The gap that bit us: an agent rewrote 400 product descriptions overnight and nobody could approve or roll back before they went live. Governance without a human approval gate and a one-click revert is just a nicer audit trail for damage you already shipped.
governance toolkit on top of a stack i still review by hand every time
Vendors keep shipping "governance" dashboards that log every agent action, but the missing layer is the human review queue, the moment a procurement bot's $40k order needs a person to approve and nobody defined who that person is or how fast they answer.
Governance toolkits assume you can log and audit what the agent did, but in a classroom the thing that needs governing is what the AI says to a 14-year-old in real time, not a compliance trail read weeks later. I've watched a chatbot confidently walk a kid through a wrong factoring method while every dashboard showed green. Audit logs catch policy violations; they don't catch being plausibly wrong.
Governance dashboards keep tracking the wrong unit: 40 agents deployed, 200 tasks automated, but nobody logs the handoff moment when a human silently rubber-stamps an agent's output without reading it. That deference rate is where the real risk sits, and no toolkit I have seen instruments it.