← Claude Code
AI Jason
What I learnt after running loops for 1 month???
⏱ 14 min video · 3 min read13 Jul 2026Worth watching
TL;DR
AI Jason shares a month of lessons from running autonomous agent loops at his company Super Design, covering the exact anatomy of a working loop: a loop contract (markdown file with goals, boundaries, SOPs), a state/log layer, trigger types, and a verifier agent. The video is a practical blueprint for building agent loops that run safely without constant human oversight.
Key points
1
Every loop needs a loop contract markdown file containing: a goal (what winning looks like), boundaries (what the agent can do autonomously vs. what needs human approval), and SOPs (step-by-step workflows to follow every run).
2
A state and log layer is critical — state holds current hypotheses and backlogs while logs are append-only run records. Without this, agents rediscover the same errors and waste tokens every session.
3
There are four trigger types: continuous for-loops (go command in Claude Code/Codex), cron jobs (scheduled runs), event-based webhooks (requires a local daemon), and combo/workflow triggers that programmatically check for new work before waking the agent — the most cost-efficient approach.
4
For high-stakes tasks, use an orchestrator-executor-verifier pattern: the orchestrator plans, sub-agents execute in isolated worktrees in parallel, and a verifier tests and attaches evidence (screenshots, video via Playwright) to PRs for human review.
5
An evolve loop should be triggered every 5-10 runs: a dedicated session where the agent reviews past logs, sharpens SOPs, cleans outdated state, and improves trigger scripts — this is how loops get better over time automatically.
Actionable insights
→
Start every loop by creating a single markdown file with three sections: goal/finish line, boundaries (auto-do vs. escalate), and SOPs. Add a state section for current hypotheses and an append-only log section. This single file is the agent's memory and constitution.
→
Choose your trigger type based on use case: use continuous for-loops for well-scoped tasks with fast feedback (bug fixing), cron jobs for daily maintenance tasks, event-based webhooks for incidents or new emails, and combo triggers (script checks first, then wakes agent) for inbox triage to cut token costs significantly.
→
Add a rule like 'never rewrite accurate documentation just to look busy' to your loop contract — without explicit anti-busy-work constraints, agents will default to producing output even when no real work is needed.
→
Set up a verifier agent that records Playwright video or screenshot evidence and attaches it to PRs. Use sandboxed environments like Daytona to run parallel verification without being limited by local dev server capacity.
→
Schedule an evolve session every 5-10 loop runs by feeding the agent the full loop config, past logs, and raw conversation history. Let it propose improvements to the contract, trigger scripts, and state — this compounds loop quality over time.
Notable quotes
“Everyone can just put together a loop using Claude Code or Codex and call it done. But that's easy 5%. The real work is how do you design guardrails that let you walk away from it.”
“Without this block, every morning the loop just rediscovers the same noise, wastes tokens on tracing the scene they had already tried.”
“One default behavior we saw agents have is that they will tend to do something even though it is not necessary — so this kind of rule is actually quite useful to making sure it produces useful results.”
Worth watching?
✅
Worth watching the full video?
Worth watching if you want to see the actual dashboard UI, live loop contract files, and the open-source tool demo — all the core frameworks are captured here, but the visual walkthrough of real running loops adds useful context.
Saved you some time? The creator still deserves a like.
Found this by searching? That's the hard way.
summree watches the AI channels you follow and sends you the summary the moment a video drops. Nothing gets past you. From £4/month.
Try it free