A shared context layer for multi-agent teams

A multi-agent context layer makes relevant knowledge available to more than one AI agent. It helps a lesson survive the session in which it was learned and become useful in another agent's work.

Your Claude Code agent might help establish why a contribution flow uses prose rather than a form. Your Cursor agent needs that reasoning when it designs a related feature next week. With Firmament, the first agent submits the lesson and the second asks for guidance from the shared knowledge base. Each agent accesses only the knowledge its user can see.

How shared knowledge moves between agents

Agent A discovers a lesson → submits it → Firmament processes it → the project is shared when appropriate → Agent B asks for relevant guidance.

The arrow describes knowledge transfer, not a live broadcast into every running conversation. A second agent needs to retrieve relevant guidance, and a submission may need processing or review before it becomes available.

This is the knowledge-sharing part of agent multiplayer. For the wider concept, read what an agent context layer does.

A design decision that informs another task

Firmament retained why its knowledge contribution process asks agents for prose with clear guidance. Structured fields encouraged box-filling; removing guidance lost the reasoning. The choice preserved room to explain what happened while asking for the decision, rejected alternatives and evidence.

A later agent designing another contribution flow can consult that reasoning before choosing an interface. Read the recorded decision and an illustrative follow-on task. The example shows how earlier learning can inform a new design; it does not claim a measured improvement in task success.

Different tools, the same project knowledge

Connect each supported agent client to Firmament. The shared knowledge lives outside the individual conversation, so opening a new session does not require manually copying the previous chat.

See how Firmament works with Claude Code and Cursor. Authentication and the user's access determine what each connection can retrieve.

Sharing is an explicit choice

Personal knowledge is not automatically visible to every colleague. Share the relevant project with a team when the knowledge should be available to its members. An agent authenticates as its user and works within that user's access.

A two-agent test under the same account verifies reuse across sessions or tools. It does not prove team sharing. To check the latter, use a second authorized teammate and verify the project's sharing settings. Read the governance page before choosing what to share.

Keeping guidance current

Agents submit new evidence as work changes. Firmament processes it into wiki pages, and people can inspect the pages and their history. New outcomes can update earlier guidance; a successful write does not prove that a claim is true forever.

Include why a procedure exists and how it was verified. If a detail changes frequently, retain the location of its source and read the current value there. Check how the wiki works for the contribution and review model.

What shared context does not coordinate

A knowledge layer is not an agent scheduler, a shared terminal or a lock around source files. It does not by itself assign tasks, stop two agents editing the same file, or synchronize the full internal state of running conversations.

Keep your existing workflow controls for those jobs. Use shared context to make relevant decisions and lessons available when each agent needs them.

How to evaluate it on your own work

Choose a task whose outcome you can check. Record what agent A learned, confirm the resulting page, then start agent B without the first conversation. Ask B for guidance using the project and task details rather than giving it the answer.

Record whether it retrieves the lesson, whether it applies it correctly, and whether the result passes the same check. A single demonstration establishes that a workflow is possible; it does not establish a general improvement in task success, latency or cost.

For a performance comparison, hold task inputs, models and scoring constant, include failures and processing costs, and repeat across representative tasks. Keep claims scoped to what was actually measured.

Put earlier lessons to use

Start with the fresh-session evaluation guide, then explore agent multiplayer to see how Firmament shares knowledge across the coding agents your team already uses.