multi-agent project coordination

Multi-agent project coordination becomes a real problem the day you run Claude Code on one task and Codex CLI on another against the same repository. Each agent is competent alone. Together they have no shared picture. Both may decide to refactor the same module, one may undo the other's change, and neither knows which task is blocked on which. The coordination layer you need is not clever. It is a single list of work with explicit scope per item, dependency information so that work is ordered sensibly, a rule about which agent may touch what, and a place where every outcome is recorded before anyone merges it. It also has to cope with a second human, since a team adds reviewers and contributors whose permissions differ. This page describes the practical pieces and where Liquid Brain stands, including the parts still under validation.

The failure modes of parallel agents

Three problems recur. The first is overlap: two sessions edit the same files because nothing declared who owned them, producing conflicts or, worse, a quiet overwrite that only shows up later. The second is ordering: a task that depends on another starts too early and builds on code that is about to change. The third is the review bottleneck. Agents produce output faster than one person can inspect it, and a board with ten items in progress simply means ten unreviewed diffs. The countermeasures are unglamorous. Give each task an explicit scope so overlap is visible before it happens. Record dependencies so ordering is data, not memory. Limit work in progress to what you can review. And keep a history of which agent did what, so when something is wrong you can trace it to an attempt rather than guessing.

Claude and Codex on the same project

Using more than one agent can be sensible: you might prefer one for exploratory reading and another for mechanical edits, or you might want an independent second opinion on a change. What matters is that both are given work the same way, from the same record, under the same rules. The task description, scope and completion criteria should not depend on which CLI picks it up. The review criteria should not either. Treat the agent as a replaceable worker and the card as the durable thing. Be cautious about letting one agent review another's work without a human in the loop, because correlated blind spots are real, and an agent agreeing with an agent is not independent evidence. Two agents can expose different problems, but the accept decision still belongs to a person.

What Liquid Brain provides, and what is still pending

In Liquid Brain, a card can open a Claude or Codex session [C1b], so both agents work from the same record, which carries scope, dependencies and completion criteria [C1a] with assignee, reviewer and blocker details [N03]. A project WIP limit can cap parallel work [N08]. Scope controls check work against declared boundaries [C5], and proposed scope stays separate from approved execution authority [N09]. Recorded attempts, outcomes and usage stay with the work [C4], and reports show recorded work history grouped by actor, though declared actor tags are not authenticated identities [C7]. The development candidate also includes project scoped collaboration, invitations, revocation and contribution review [C8]. Those controls remain under validation: distinct denied identity and host isolation checks are still pending, so this is not a production multi user claim.

QUESTIONS

Can I run Claude Code and Codex CLI on the same project?

Yes, as separate tools, if each is installed and configured. The coordination problem is overlap and ordering, so give each task explicit scope and dependencies. Liquid Brain opens either agent from a card [C1b], working from the same record [C1a].

How do I stop two AI agents editing the same files?

Declare scope per task, record dependencies, limit work in progress and review before merging. Liquid Brain stores scope and dependencies on the card [C1a], supports a WIP limit [N08] and checks work against declared boundaries [C5]. None of this is a universal sandbox.

Does Liquid Brain support team collaboration?

The development candidate includes project scoped collaboration, invitations, revocation and contribution review [C8]. These controls are under validation, though, and distinct denied identity and host isolation checks are still pending. Treat it as not yet proven for production team use, and do not assume a release exists.

RELATED GUIDES

• What is a project management harness?
• Cards vs scheduled tasks: what starts an agent?
• How to find and organize work across projects

RELATED HARNESS PAGES

• Claude Code project management Claude Code project management as a workflow: run Claude from project cards, schedule tasks and review results before you accept them.
• Codex CLI project management Codex CLI project management as a workflow: launch Codex from project cards, schedule explicit tasks and inspect outcomes before accepting.
• kanban for AI agents Kanban for AI agents: the board states that matter when software does the work, including review, blocked and decision owed.