Agent platform of 2027? This open-source project wants AI agents on the same board
Jiayuan (JY) Zhang has released the open-source project Multica, a collaboration platform that puts coding agents such as Claude Code and Codex on the same board. John Suh calls it the Agent platform of 2027 and the endgame at the application layer. The project aims to solve the chaos of one person managing a pile of agents and the blank space in multi-person, multi-agent collaboration.

Multica is released on GitHub with a one-line description: "Agents that show up on the board." It positions itself as a source-available workspace that treats AI coding agents as teammates and assigns them work: agents pick up issues themselves, report progress, raise blockers, and hand work back for review. It is self-hostable, compatible with the agent CLIs you already use, and does not lock you in.
After the tweet went out, John Suh reposted it and gave it a bold characterization: "Agent platform of 2027. This is the end game." He further explained that this would be the only application-layer software that exists, and that all other software would become data sources for this layer. This is the application-layer endgame for ChatGPT and Claude. He also said the right agent platform should be like a garden, where customers plant their own seeds, rather than a "factory of factories." These are very sweeping claims, but as a starting point for discussion, they can make people re-examine the direction of agent platforms.
The practical problem Multica wants to solve is very concrete: you are already running Claude Code, Codex, and three other agents, each living in its own terminal tab, forgotten as soon as the session ends, and today you are explaining the same context for the fourth time. The more agents you have, the more time you spend "babysitting" them. Multica puts these agents and your teammates in the same workspace, assigns an issue, and the agent handles it on a runtime you control, comments as it works, and finally hands it back for review. Intent, execution process, decisions, and diff all hang under the same issue; there is no such thing as rebuilding context, and no code goes live without human confirmation.

In terms of features, it supports 31 kinds of agent CLI, from Claude Code and Codex to Cursor, Kimi, and others. You can give each agent a name, provider, and runtime, and they appear on the board like ordinary members. It also has autopilots, chat, project grouping, and more. Every run has a complete execution log, you can replay every tool call and command, and it tracks the cost of each agent. Work can only enter review status, and a human decides whether it can be merged. The inbox only notifies you when you need to make a call, instead of interrupting you at every step.

As for how it differs from similar products, the author gives a concise layering:
- Personal + single agent: Claude Code/Codex/DSH/Pi
- Personal + multiple agents: Orca/T3Code/Herdr
- Multiple people + multiple agents: Multica
In other words, Multica targets organization-level multi-agent collaboration, but the author also says it is a superset of the first two and works fine for personal use. One commenter summarized it more vividly: Herdr gives every worker a workstation, T3 Code is you sitting next to them directing via walkie-talkie, and Multica is a dispatch system where work orders come in and are automatically assigned to available workers, the process is recorded, results are accepted, and multiple people share it. Another commenter said, "So the boss can use the Claude Code that employees bought through Multica," and the author replied: indeed.

Architecturally, the backend is written in Go, the database is PostgreSQL 17, and the agent daemon runs on your machine and connects over WebSocket; it is responsible for launching CLIs such as Claude Code and Codex. It can be self-hosted with Docker Compose or Helm, and a cloud service is also offered. The self-hosted version sends an anonymous deployment-level snapshot every day, containing no content, and it can be disabled by setting DO_NOT_TRACK=1. Note, however, that it does not include models and does not have built-in agents; it only drives CLIs that you have already installed and logged in to.
Someone asked whether it is similar to Vesence; in fact, the two point in somewhat different directions. Some commenters also worried about data leakage, and the author asked why in response. This is a common privacy concern; it depends on how you run it.

Back to John Suh's assertion: if the agent platform really is, as he says, the only application-layer software that survives, then projects like Multica are at least staking out that position. Even if his prediction is too aggressive, the chaos of multi-agent collaboration is real, and managing agents as colleagues is a concrete answer worth trying.
发布时间: 2026-11-23 05:40