Don't trust an agent that says "done."

Peter Steinberger—OpenClaw’s author—posted on Twitter a couple of days ago. It blew up:

Peter Steinberger's post on X — Peter’s post

After heavy use of Codex and Claude Code’s Goal loops—and falling into the same holes N times—I had to feel out a next-generation Agent loop framework.

Flow was designed around a graph structure from day one. It sidesteps the open-loop problem: when the model decides what to do next and when the work is done, complex tasks flip the car fast.

Open loop vs Flow supervised loop — Open loop vs Flow

Plenty of us have used agents on projects. A Goal loop feels fine on simple tasks; with a large codebase or complex work, you start to feel the cracks.

You say what you want and the model starts grinding. It doesn’t push you to clarify, and it never really gets what you expect, so it guesses from similar patterns in its weights.

The result disappoints: it says done with incomplete features or junk quality. Built-in Codex / Claude Code review is hit-or-miss too.

A few reasons.

Why agent reviews fail: executing before alignment, player and referee, review lacks rigor — Three common process risks

One: you and the model never fully aligned, but vendors want agents that look end-to-end smart, so the harness pushes confident execution (see Codex system prompts). Alignment gets skipped unless you force plan mode or stack skills.

Two: coding-agent review design is shaky. Some let the same model judge its own work, so fairness is gone. Others only look at git diff, not this session’s edits—so messy habits (mine included) pull in parallel work or old changes.

Three: review isn’t multi-model adversarial, and context isn’t separated, so the check isn’t strict, objective, or complete enough.

Flow’s bets, short:

Grilling prompts until intent is actually pinned.
Context extraction as evidence, labeled as user↔executor dialogue—reviewers stay read-only + tests.
On failed checks, especially even rounds, a stronger advisor redirects—plan and execution, not vibes.
Prewalk: strong-model planning context injects into executors so they re-read less and wander less. Source: stencil.so/blog/prewalk.

Live HTML reports—stop staring at a terminal wall. LAN or Tailscale on your phone. Archive later.

Install on a machine with Pi: pi install npm:@surgee/pi-flow, set role models, /flow.

My usual mix: advisor Fable / Kimi K3, executor GPT 5.6 sol Xhigh, reviewers GPT 5.6 sol high + Terra xhigh + Sonnet 5 high.

Enjoy. Build everything in the loop. Ship the task, then go get coffee.