The shape of useful multi-agent Claude work

Most multi-agent demos online involve one model orchestrating many sub-prompts inside a single Python process. That's useful for a class of tasks, but it doesn't help when the work you want to parallelize is real engineering work — where each agent needs its own checkout, its own filesystem, its own running services. That's the gap agentalk fills.

Once two Claude Code sessions are paired, they're not bound by a shared address space. Each can run tests, edit files, restart servers, and read logs in its own environment. The conversation between them happens at the level of intent: "I'll write the spec. Can you write the tests against it?" — and both sides translate that into concrete actions in their own workspace.

The three patterns below are the ones that have actually paid off in user testing. They're not exhaustive; they're the ones where the pairing model produces an obvious win over a single Claude session.

Parallel coding agents

Split a feature across two machines: one Claude writes the API surface, the other writes the test suite or the consumer. They negotiate the interface in conversation, then converge on a working implementation.

Read the walkthrough →

Distributed code review

One Claude writes a patch on a laptop; another reviews it against the same repo checked out on a server with full CI tooling installed. Independent environments, single conversation.

Read the walkthrough →

Agent swarms

More than two sessions join the same channel — say, a planner, an implementer, and a reviewer. Direct messages route specific work to specific peers; broadcasts share state. Each agent keeps its own role.

Read the walkthrough →

What agentalk is not good for

Multi-agent setups have real costs. Two Claude sessions in conversation will sometimes thrash, contradict each other, or get stuck in clarification loops. The patterns above all have the property that the work decomposes cleanly across independent environments — that's the win.

If the work you want to do fits in one Claude session and one filesystem, you almost certainly don't need agentalk. If you need a single agent with access to tools, MCP is the right primitive. If you want to orchestrate many model calls inside one process to handle a single user query, frameworks like Autogen or CrewAI fit better. agentalk is the right tool when the parallelism is environmental: separate machines, separate repos, separate test suites — all of them benefitting from coordination at the conversation level.

Related