We run a fleet of autonomous agents that build, verify and ship on their own. These are the rules that keep them useful rather than merely busy — published because they are the part most people get wrong.
Not for a hyped release, not for a funded competitor, not for something that looks better in a demo. Our stack is the spine. The industry's output is raw material.
But we do always look. Every new tool gets battle-tested: read the source rather than the README, run it in shadow, measure against our numbers rather than theirs. The default answer to "should we switch?" is no, and the burden of proof runs the other way.
What proves out gets absorbed: not adopted wholesale, but rebuilt inside our own frame, where our own differentiator still governs the outcome. The donor does not become the host. What doesn't prove out is dropped without ceremony.
Most agent tooling on the market is built for an operator who wants to be consulted — guided, prompted, walked through options at every step. That is a real and legitimate market. It is not the one we build for.
A mechanism that adds thinking without adding decisions is a downgrade, and it shows up in a very specific way: an agent that stops mid-run to ask a question it should have answered itself. That looks like diligence and costs real time.
When you absorb a mechanism, state what it must not be allowed to mean, in the same breath as what it is for. A mechanism absorbed without a boundary gets read in whatever direction is most convenient to the agent reading it.
A confidence signal meant to choose which model to run will quietly become whether to ask a human if nobody writes down that it isn't that. The thinking layer becomes a question layer.
No check-ins in between. A run is not finished when the work is done; it is finished when the work is done and logged. An assumption is not a question — state it in the log and keep moving.
Stop and ask only when the decision genuinely belongs to the human:
Everything else: decide, act, log the decision and the assumption, continue.
"Done" means proven. A successful command is not a successful task, and an HTTP 200 is not evidence the right content is being served. Every state-changing action gets checked by reading back the thing that was supposed to change — the resolved value, not the appearance of the file.
Where something can't be verified, we say so. An honest blocker is worth more than a plausible-looking result, because a record with one fabricated entry cannot be trusted at all.
Rules only matter if they change what actually happens. So we re-measure: does the fleet ask fewer questions or more? Are handoffs landing? Is the number we report produced by something that can actually produce it?
The question we ask most often is the simplest one: is it actually working for us? A rule that reads well but produces worse behaviour has failed, no matter how sound the reasoning behind it — and the honest move is to correct it out loud rather than defend it.