Pimp My IDE / Garage dispatch
Back to garage
September 27, 2026 | coding agents / coordination / review

A team needs a handoff, not a headcount.

More agent seats can produce more work. They can also produce conflicts, duplicated judgment, and a patch nobody owns.

The take. Name each seat, give one seat patch ownership, preserve an independent objection route, and make every reviewer inspect the exact candidate that may ship.
Set the handoff clutchesInspect OpenRig

The rig is real.

OpenRig packages a local daemon, command-line interface, terminal interface, MCP server, and tmux-based runtime adapters. Its current repository documents persistent seats, named addresses, messages, queues, snapshots, and mixed Claude Code and Codex rigs.[1]

The install artifact also exists. The garage fetched the public npm metadata for @openrig/cli version 0.5.17 and unpacked its tarball on September 27. The package declares Node.js 20 or newer and exposes the rig command. That proves there is a package to inspect. It does not prove the package is safe for your machine or that a multi-agent run will improve your code.

A topology tells you who can talk. A handoff tells you whose work can ship.

Parallel work has a shape.

OpenAI's Codex documentation recommends parallel agents first for read-heavy work such as exploration, tests, triage, and summarization. It warns that parallel write-heavy work can create conflicts and coordination overhead.[2]

Anthropic documents a similar separation. A Claude Code subagent gets its own context, prompt, tools, and permissions. Anthropic also distinguishes one-session subagents from independent sessions and coordinated agent teams.[3]

Those facts do not make one orchestration product correct. They do expose the contract that any harness must make visible: work division, authority, review independence, and the artifact under review.

Keep one patch owner.

Let many seats search, test, or challenge assumptions. Give one seat the candidate patch. This removes the easiest coordination bug, where two writers edit the same area and the final diff contains an accidental compromise.

The checker should receive the commit, branch, or patch identifier. "Review what the builder did" is not enough. A persistent seat name is useful, but only an exact candidate binds the review to the code that may ship.

Wire dissent into the route.

A reviewer who reports through the builder can become a second voice for the same conclusion. Give the checker a direct stop path to the merge owner. Record unresolved objections beside the candidate, not in a terminal pane that disappears.

The differential below turns four decisions into a copyable handoff template. It does not start agents or inspect a repository. It makes the route explicit before the run begins.

Interactive makeover / agent handoff

Handoff differential

Instead of arranging agent windows and inferring ownership, select four handoff sections, see the route connect in order, and copy a packet that still requires real names, candidate IDs, and review results.

Engage the route

The shaft connects only through a contiguous prefix. A later clutch can be selected while an earlier gap keeps the route blocked.

IdentityName the lead, builder, checker, and merge owner.
Patch ownerOne seat writes the candidate. Other seats return findings or tests.
DissentThe checker can report a block directly to the merge owner.
CandidateRecord the exact commit, branch, or patch that received review.

0 of 4 handoff sections selected. Route blocked at Identity.

Handoff template0 SECTIONS SELECTED

Copy the run contract

Selections add required fields to the packet. They do not fill those fields or prove a review happened.

What completion means. Four sections are present in the template. Real seat names, candidate identifiers, review findings, and merge evidence are still required.

Sources and limits

Open the source log
  1. OpenRig repository and README, read September 27, 2026. It documents named seats, mixed Claude Code and Codex rigs, tmux sessions, queues, messages, snapshots, startup writes, permissions, and the @openrig/cli package. The garage also fetched npm metadata for version 0.5.17 and unpacked the public tarball. It did not launch a rig or grant the package workspace access.
  2. OpenAI, "Subagents", read September 27, 2026. It documents parallel delegation, per-agent model settings, token costs, inspectable threads, and the conflict risk of parallel write-heavy work.
  3. Anthropic, "Create custom subagents", read September 27, 2026. It documents separate context windows, prompts, tools, permissions, usage accounting, and the distinction among subagents, independent sessions, and agent teams.
  4. GitHub Trending repositories, read September 27, 2026. It surfaced OpenRig during this editorial pass. Trending placement is a discovery signal, not evidence that the tool works or is safe.

Source boundary. The product claims come from the OpenRig project. The operating advice is the garage's synthesis of that project and the two vendors' agent documentation.