Let agents act, inside guardrails
Agents that act as well as suggest: autonomy tiered by risk, with a rollback path and an audit trail. Aimed first at the work that should not exist.
The problem
When AI only suggests and a person still types every change, an organization gets a small part of what these tools can do. Banning the tools outright gives up most of the rest, and the safety it seems to buy is not real: the risks are still there, and now nobody is watching them. Most of the value comes when an agent is allowed to act on its own, for example to make the change, run the tests and open the review. That works only inside a system that limits what the agent can do, records what it did, and lets a person undo it.
How I approach it
- Autonomy tiered by risk. Safe, mechanical changes run on their own. For anything higher-risk the agent classifies and proposes while a person runs it.
- Rollback and audit. Every automated action has a rollback path and an audit trail, so “it acted on its own” never means nobody can see what it did or undo it.
- Guardrails the organization owns. The organization checks what goes into its agents and what comes out of them: the data and instructions an agent is given, and the changes, messages and answers it produces. Its own people write those checks and run them. A vendor’s monitoring is built for its own product and for all of its customers, and it does not know the organization’s rules, its data or what counts as a mistake there. The rules are proportionate to the real risk.
- Scripts for calculations. When an agent needs a number, it writes a script that calculates it instead of doing the math itself. A script is deterministic and an LLM is not: the same input always gives the same number, which matters wherever the number is money or a metric someone will act on.
- Rules the agents write. When an agent can identify the check that would have caught its mistake, it writes that check down as a dated rule, and the rule is dropped when its reason stops being true.
- A human yes. Nothing goes outside the organization without a person’s yes.
- Tools for repeated work. The third time a job is done by hand, someone builds an instrument for it.
- Aimed at the work that should not exist. The first place I aim agents is an organization’s own repetitive operational work. The strongest version does not do that work faster: it prevents the work from being created at all.
- A date on every bet. Every AI bet has a date, agreed before it starts, when it is either stopped or scaled up, depending on what it produced by then.
- No quotas on usage. I do not set quotas on how much people use the tools. A quota rewards usage, so people use the tools to hit the number. What I want to reward is results.
What it produces
Agents doing real work under rules the organization can read, audit and undo.
In the Luci House
The AI-staffed team I run every day as a working prototype.
The members act on their own inside gates they cannot open. A change is stamped verified only when the full suite passes, the deploy step refuses anything unverified, and the production copy of the code is owned by a separate account, so nothing can edit the running system in place. The members also write their own rules: on November 20, 2026, the seven members of the working team carried 138 rules they had written for themselves, 66 of them written in the previous 30 days, and a rule is dropped when its reason stops being true. When the same job is done by hand for the third time, the expectation is to build an instrument for it. How the members learn
Figures as of November 20, 2026.