Built in public · AI engineering

An unattended agent fleet,
held to evidence.

I design and run a multi-lane autonomous agent system on one machine. Each lane researches, builds and ships small products with no human in the loop. The products are the testbed. The work is the engineering that makes unsupervised agents reliable: verifiable claims, mechanical guards, model tiering, and failure handling that doesn't depend on the model behaving.

What it shipped

Small digital products that the lanes research, build, list and promote themselves. Counted from the fleet's own ledgers and re-checked on every publish. Revenue appears here only once a scripted read of the storefront confirms it.

Architecture

One hourly runner. Deterministic gates run before any model is called, and a verifier checks everything a model wrote before it's committed.

Planning · daily

Planner

Sonnet daily, Opus weekly review. No tools; reads outcomes, emits playbook ops.

Shared playbook

ADD / EDIT / RETIRE / PROPOSE ops, append-only change log.

Focus cap

Five focus areas at most; the rotation follows them.

Execution · hourly, per lane

Preflight gates

Pause flag, network, live usage windows, daily run cap.

Script reads

Metrics are read by code first, so the model can't invent them.

Stall rule

Three no-progress runs on one item park the lane.

Worker run

Headless agent, restricted permission profile, hard timeout.

Verifier

Stage claims without stage-typed verified_via go to quarantine.

Commit

Every run leaves a commit and a log; blockers go to the owner.

Side channels

Volume tier

Free-model chain behind a local proxy for bulk sourcing and scans.

Paced publisher

Deterministic script with cap, gap and URL check, outside the LLM loop.

Protected paths

Guards and runner are read-only to agents; edits are reverted.

Model callGate (can stop the run)Deterministic code

Runs per week

CompletedNo output or error
Table view

Decision log

How the fleet was steered: what the signal was, what changed, and what that bought.

    Real examples

    Raw excerpts from the fleet's state files, verifier source and git history. Picked by hand; product, platform and lane names are replaced with [brackets]. Everything else is as the fleet wrote it.

    Principles that held up

    Capability, not judgement

    If an agent must not do something unattended, remove the permission. Don't ask it nicely.

    Claims carry their proof

    A status is a pointer to a live check, or it's quarantined.

    Deterministic where possible

    Reading numbers, pacing and publishing are scripts. The model plans and builds.

    Fail loud, once

    A stuck lane parks and files one blocker, instead of burning runs in silence.