Everyone is building agent frameworks right now. I went the other direction — the most reliable parts of my AI-agent setup are the ones with zero model involvement.
I run side projects where agents (Devin, Codex, T3 Code, whatever) do most of the work. The agents were never the bottleneck. The missing pieces were operational: sessions forgot state, workers claimed things that weren't true, deterministic work burned model tokens, and human-gated steps stalled because I wasn't at the terminal.
Here's the layer I actually run — all of it is boring, and that's the point.
Deterministic first: the model only does judgment work
A GitHub Actions cron scans for newly funded bounty issues daily: gh api search, dedup against a seen-file, commits a report, opens an issue on finds. Cost: $0/day. No model call could do this better — it's a filter, not a judgment.
Rule of thumb that fell out of this: if the task has a spec, it's a script. The model gets routed only the work that genuinely needs reasoning.
A state file, not a memory
AI_HANDOVER.md at repo root: verified current state only, rewritten each milestone. Every session starts by reading it, and the fix for "every session starts from zero" turned out to be a file, not a vector DB. Two rules make it work: only verified state goes in (checks are cited, not asserted), and it's rewritten rather than appended, so it can't rot.
An append-only decision log as the audit trail
Any irreversible action — a publish, a payout, a post — needs a logged approval in DECISIONS.md first. It's a markdown file with one line per approval. When something goes sideways, the question "who approved this and why" has a grep-able answer.
Approvals from my phone, not my desk
tg-post.sh "msg" pushes to a Telegram bot; tg-poll.sh long-polls the inbox into a log file. A launchd plist keeps the listener alive across reboots and works when the agent app is closed. ~30 lines of bash total. Slack is the same shape with a different endpoint. The effect: an agent can ask "publish this?" at 2am and I approve from my phone — the human gate without the human having to be at the keyboard.
Delegation contracts with verification gates
Every delegated task gets a WORKER_BRIEF: scope, files, acceptance checks. The lead agent re-runs the checks before the work counts. Workers reporting "done" is a claim; the lead re-running the test is the evidence. This single habit killed most of my "looks done, isn't" failures.
What it is not
It's markdown + bash + one yaml file. No framework, no SDK, no lock-in. Runner-agnostic — the same files work under Devin, Codex, Claude Code, or cron alone. Secrets live in a chmod-600 env file outside the repo; scripts take env vars, never literals.
I packaged the whole layer — the doc templates, comms scripts, the bounty-radar workflow, worker-brief format — as a starter kit for people who want the files: agent-ops-starter-kit on GitHub (free samples in samples/; full kit is a paid download, $39–149).
But honestly, the two patterns worth stealing even if you never buy anything:
- Verified-state handover file — fixes session amnesia for $0.
- Worker-brief + re-run-the-checks gate — fixes worker overclaiming, same price.
Happy to go deeper on either in the comments.
