The patterns

Time

5 minutes to read.

You need

• Loopgate open: On your laptop

Two harnesses on one problem

• parallel-plans-then-build. Two plans compared before any code. Planners on Claude Code and Codex plan alone; a lead writes one PLAN.md from both, you lock it, and a builder builds it under review for up to two rounds.
• root-cause-fix. A cause backed by two independent looks. Read-only investigators on each harness look alone; a lead writes one RCA.md; you fix it under review for up to two rounds, or finish with the report.
• pair-programming. A second model watching every step. A driver on Claude Code works in small steps and a navigator on Codex reviews each one, for up to six rounds.
• adviser. Help when the worker is stuck. One agent does the work and asks a read-only adviser on Codex, for up to four rounds.

Build, check, review

• one-agent. The simplest run. One agent with Workspace access that waits for your reply after each turn.
• plan-build-review. A plan you lock and a reviewed change. A planner writes PLAN.md for you to lock, a builder builds, and a read-only reviewer approves or sends changes to a fixer, for up to two rounds.
• architect-build-test. Behaviour proven by tests. An architect's design you lock, a build it revisits for up to three rounds, then a tester and a fixer for up to three more.
• build-fix (needs a declared verify). A command, not a model, says when it is done. A builder, a check that runs your verify command, and a fixer while it fails, for up to three rounds.
• reviewer-panel. Two reviews of one change. A builder builds; security and quality reviewers read the diff at once; a fixer fixes what they found.
• security-audit. A report first, fixes only where you choose. An audit step runs Cloudflare's security-audit skill; you choose what to fix; a fixer fixes and a verifier with the same skill checks, for up to two rounds.

Work split in parallel

• per-module-builders. Two modules at once. A plan you lock, two builders on the API and the UI in their own worktrees, then a reviewer and a fixer on the merged whole for up to two rounds.
• work-items. Your own list, built at once. Write ## Work items in the task with a ### section per item; a builder takes each, three at a time, and a lead verifies the merged work.
• orchestrator-items. A lead who splits the work and checks it. Builders take the lead's work items, and the lead reviews the merged result with a fixer for up to three rounds.
• orchestrator-team. A fixed team under a lead. The lead assigns the API and the UI to two builders, verifies their merged work and reassigns, for up to four rounds.
• orchestrator-lanes. One conversation that plans, delegates and verifies. An orchestrator plans with you, writes a brief per lane, runs each on the model it names and hears every lane, for up to six rounds.
• design-build-panel (needs a declared verify). Ticket to merge in one run. An architect and a critic agree a design, a builder per package builds, a lead integrates and verify runs with repair; seven read-only reviewers, one on Codex, read the diff for a fixer, and you approve the pull request before it opens.

Or say one

Describe the shape in the composer with Who on Auto, such as "two builders take the API and the UI at once, a reviewer reads both, then I approve the pull request". The copilot writes the workflow, draws it on the draft and names under Left out what it could not build as you said. Open in builder keeps it among yours.

The built-in agents on the Agents page work alone or as one step: Planner, Investigator, Architect, Builder, Reviewer, Fixer, Tester, Scout, Driver, Navigator, Adviser, Lead, Orchestrator and Assistant.

Steps

1. On Workflows, open plan-build-review and choose Simulate under the canvas.

   You should see: every route walked on the fake harness, the loop going round and ending at its decision, and no run created.

2. On Tasks, pick plan-build-review under Who and say a task.

   You should see: its steps under the box, each with its engine. Click one for its setup and Preview every word.

Good to know

• Each template names engines. Loopgate fits every step to an engine you are signed in to and records what it changed.
• A built-in is read-only. Make a copy to edit gives you your own: Build your own.

Next

• Build your own
• A model per step