Compare multi-agent development approaches
Compare multi-agent development approaches
This page compares workflow guarantees, not model intelligence. Codex CLI, Antigravity CLI, and other tools can
each solve many tasks alone; coordinate-agents is useful when separation of duties, role-based handoffs, and a
recoverable local bus are worth the structured process.
| Approach | Implementation owner | Review owner | State recovery | Release gate | Shared credentials |
|---|---|---|---|---|---|
| Single agent | Same agent | Same agent | Depends on the tool | Usually none | None |
| Manual dual terminals | Not fixed | Not fixed | Weak or manual notes | Manual | None |
coordinate-agents (Protocol & Runtime) |
Configured Implementer (antigravity by default) |
Configured Reviewer (codex by default) |
Local persistent .agent-bus |
Exact user authorization | Not required |
What the differences mean
Single agent. This has the lowest coordination cost and is usually best for small, reversible changes. The same context plans, edits, tests, and judges its own output. Independent review and a durable cross-tool handoff are not provided by this project.
Manual dual terminals. Two CLIs can share a repository and communicate through chat or ad hoc
Markdown files. This is flexible, but role ownership, message claiming, interruption recovery,
deduplication, and release authorization depend on both prompts being followed perfectly. Two
processes can accidentally modify the same worktree. Codex App can remove the manual terminal
overhead when the Skill is invoked from the correct project path, but the configured Implementer
still runs in a Runtime-owned persistent Session and must use a real executable command such as
agy or claude.
coordinate-agents. The coordination protocol and runtime decouple agents into configurable workflow roles (planner, implementer, reviewer) and independent project/Agent-scoped Execution Sessions. In the default reference workflow, Codex clarifies requirements, writes acceptance criteria, reviews the implementation commit and evidence, and controls the release gate, while Antigravity is the exclusive product-code writer. Messages move atomically through new, processing, and processed; leases, bounded PTY I/O, Session reuse, and recovery preserve work across terminal exits. Neither agent reads or copies the other’s authentication files. Dynamic registration allows connecting third-party agents (generic-cli) with customized role assignments.
Observable guarantees and limits
The bus records who sent a message, its type, claim state, related commit, and evidence path. The
deterministic demo verifies IMPLEMENT -> IMPLEMENTATION_DONE -> REVIEW_APPROVED on the default reference workflow and reruns tests.
The installer refuses unrecognized or modified directories unless the user explicitly forces a
destructive action.
These are workflow controls, not a security sandbox. .agent-bus is local plaintext, any process
with filesystem access may read it, and the tool cannot prevent a deliberately disobedient CLI from
editing files. Git history, tests, OS permissions, and human release authorization remain necessary.
Do not use this project when
- one agent can complete the task safely and independent review is unnecessary;
- multiple agents must edit product code concurrently on the same worktree without worktrees;
- agents work in different repositories or on machines without a shared filesystem;
- the project requires encrypted or remotely replicated coordination state;
- an organization needs mandatory identity, policy, or cryptographic approval enforcement;
- the repository cannot tolerate a local ignored
.agent-bus/directory.
Use separate Git worktrees or a different orchestrator when concurrent implementation by multiple writers on the same worktree is the actual requirement. For the default workflow’s exact ownership rules, read the Codex role (reference), Antigravity role (reference), protocol specification, and security boundary.