OpenClaw splits QA testing across 12 coding-agent subagents

• OpenClaw's release prep moved QA into a long-running Codex run: steipete's prompt assigned 12 subagents, live API-key E2E tests, worktrees, PRs, and a target of 200 bugs.
• A Codex chief-of-staff loop checked Slack, email, and Linear at 9 am, 1 pm, and 5 pm; the heartbeat clip says the loop later added links, Linear context, and pre-drafted replies.
• The strongest review pattern was model diversity plus hard gates: om_patel5's workflow used a fresh context, a different model, a severity rubric, and binary tests or schema checks.
• Skills became portable workflow memory: shannholmberg's breakdown maps 22 Matt Pocock commands across interview, planning, build, review, and handoff stages.
• Creator-facing coding moved into UI and voice surfaces: emilkowalski's prototype used variant switchers and Leva controls, while moritzkremb's recording taught a task from a narrated desktop demo.

steipete's screenshot shows an eight-hour QA run with active bug hunters, 16 completed subagents, and 55 independently verified fixes already counted in the OpenClaw QA post. OpenClaw's public agent-skills repo calls autoreview a shared closeout workflow, with Codex as the default reviewer and Claude optional. Jason Liu's Codex system turns a thread into a scheduled heartbeat across Slack, email, and Linear in petergyang's clip. The UI builders were doing the same pattern for taste: emilkowalski had agents generate variants with a query-param switcher and Leva controls in his prototype video.

steipete gave Codex a release QA target, not a narrow ticket.

The prompt assigned the agent to:

• run a full end-to-end OpenClaw QA test with live API keys
• use 12 subagents to split functionality
• spin up dev gateways on different ports
• stress test parts of the system
• orchestrate with worktrees
• create PRs autonomously
• find 200 bugs
• fix root causes, not band-aids
• avoid touching the plugin SDK boundary unless required
• keep a Markdown test report updated on the desktop

In a reply, steipete said the run used a 12-core VM steipete's VM reply. gregisenberg's cloud-agent post framed the same systems move around isolation: each agent gets its own machine, so parallel work avoids local worktree collisions gregisenberg's cloud-agent post.

Jason Liu, described by the Codex episode post as DevEx at OpenAI, used Codex as a daily operations loop rather than a code-only assistant.

His examples covered:

• reading the past week of his Slack messages and making a skill that learns how he talks
• turning a thread into a heartbeat that checks email, Slack, and Linear
• adding links after the first run omitted them
• layering in Linear context
• pre-drafting Slack replies and emails
• finding a booking ticket, checking into a flight, and texting a boarding pass

The Jason Liu episode page describes the same system as real prompts for a Slack/email chief of staff, mining past sessions for new AI skills, and running long tasks with verifiable goals.

LLMJunky argued that newer models are being prompted with outcomes, guardrails, tool boundaries, and a definition of finished, not step-by-step spec sheets.

The same reply left room for constraints: preferred frameworks or required implementation choices still belong in the outcome. OpenAI's prompt guidance says GPT-5.6 can often infer the user's goal without every step being prescribed, while still needing domain context, hard constraints, approval boundaries, and success criteria.

The OpenAI docs split enforcement from approval: OpenAI's guardrails page says guardrails validate input, output, or tool behavior, while human review pauses a run for approval decisions. thekitze compressed the anti-plan mood into one line: planning is becoming the new “looking at code” thekitze's reply.

shannholmberg broke Matt Pocock's bundle into 22 commands across six workflow buckets:

• Interview: /grill-me, /grilling, /grill-with-docs
• Domain and design: /domain-modeling, /codebase-design, /improve-codebase-architecture
• Spec and plan: /to-spec, /to-tickets, /wayfinder, /triage
• Build: /implement, /tdd, /prototype, /research
• Review and fix: /code-review, /diagnosing-bugs, /resolving-merge-conflicts
• Workflow: /ask-matt, /handoff, /teach, /writing-great-skills, /setup-matt-pocock-skills

The core pattern is grill first, then spec, build, and review. The Best of JS page for Matt Pocock's skills lists npx skills@latest add mattpocock/skills for setup and Claude Code plugin commands for managed installs.

The thread sits inside an “all of the above” builder mindset, as dustinhollywood's reply put it, but the mechanism is budget routing.

His workflow uses the highest-intelligence models for architecture and review, then cheaper models for bulk execution:

1. Use FABLE 5 ULTRACODE to architect the product, challenge assumptions, and write the implementation plan.
2. Send the plan to CODEX SOL 5.6 ULTRA for independent review.
3. Have the reviewer look for missing systems, security issues, UX/UI gaps, scalability issues, performance bottlenecks, and simplifications.
4. Return the findings to FABLE 5 and repeat until reviews stop finding meaningful issues.
5. Use Claude 4.7 or GPT 5.4 for the bulk implementation work.
6. Keep PRs small.
7. After each PR, use GPT Sol Ultra for requested changes and FABLE 5 for an independent review.

The important constraint in the post was evidence grounding: every recommendation had to trace back to code, docs, benchmarks, tests, or other verifiable facts.

om_patel5's review loop had four steps:

1. Claude writes the code.
2. A fresh context window reviews it.
3. A different model reviews it too.
4. Both reviews go back to the original Claude session, which either fixes issues or explains why the reviewers are wrong.

Two controls did the heavy lifting. The reviewer got a severity rubric so every nit did not become critical. Tests and schema checks handled binary truth, leaving model reviewers for judgment calls.

OpenClaw's public autoreview skill turns that review loop into a closeout gate. The linked skill file from steipete's post describes an advisory code-review workflow, not an automatic permission to ship.

The OpenClaw autoreview SKILL.md says:

• Codex review is the default engine.
• The default Codex model is gpt-5.6-sol with high reasoning.
• gpt-5.6-terra is the fallback when Sol is unavailable.
• Claude review is optional and uses claude-fable-5 by default.
• --panel runs Codex plus Claude unless the first reviewer is changed.
• behavior-validator covers user-visible behavior because autoreview judges the source bundle.

The file states the caveat plainly: a clean source-aware review is not proof that a UI, CLI, API, or generated artifact works from the user's perspective.

emilkowalski's prototyping loop asked the coding agent to generate multiple feature variants, add a switcher, and persist the selected variant in the URL query string.

That made a shareable design artifact: a colleague could open a specific variation by URL. The workflow paired the switcher with the Leva repo, which exposes React controls for live parameter tweaking. emilkowalski later said he used his own skill for the workflow emilkowalski's skill reply.

moritzkremb tested “record a skill” by narrating a desktop workflow. The attached UI captured a 146.3-second recorded demonstration for building a newsletter welcome sequence, including spoken context about audience mix, Beehiiv automation, and content selection.

He compared the experience to recording a Loom for the AI. In a follow-up, he said he builds several skills throughout the day and records when it fits moritzkremb's follow-up.

Voice was moving into the same delegation lane. minchoi's Codex video showed a user talking to a coding agent in real time, interrupting it, and sending it to edit files minchoi's voice-mode post.

steipete turned an agent workflow into a graph: planner, worker, independent plan reviewer, three concurrent reviewers, synthesizer, pass/fail gate, feedback, and worker retry.

The run used 12 actual model calls and two review rounds. The first attempt was deliberately made to fail, the second passed all three reviewers, and no repository files were changed.

Gavmn showed the same delegation shape in product chores: ask the agent to boot a simulator, take screenshots, add them to a PR, loop until CI is green, request a teammate review, merge when possible, and check the task off in Notes Gavmn's delegation reply.