Agents Can Overload Their Own Context - Use Separate Context with Shared Log
Agents Can Overload Their Own Context - Use Separate Context with Shared Log
The first instinct when running multiple agents is to give them shared context. Let them all see what the others are doing. Full transparency. Maximum coordination.
This kills performance.
The Shared Context Problem
When three agents share a context window, each one pollutes the others' context with irrelevant information. The frontend agent does not need the backend agent's SQL debugging output. The test agent does not need the frontend agent's CSS iterations.
But shared context means everything goes everywhere. Context fills up faster. Each agent spends attention on information that is not relevant to its task. Response quality degrades because the signal-to-noise ratio drops.
Separate Contexts, Shared Log
The pattern that works is isolation with coordination. Each agent gets its own context window containing only what is relevant to its task. Coordination happens through a shared append-only log.
The shared log is structured and minimal. Each entry contains: who did what, what changed, and what other agents should know. An agent reads the log when it needs to coordinate but is not burdened by the full context of other agents' work.
How the Log Works
The backend agent finishes the API endpoint and writes to the log: "POST /api/users endpoint ready, accepts , returns ." The frontend agent reads this single entry and knows exactly what to integrate with. It does not need the 50 messages of backend debugging that produced the endpoint.
This is information compression. The full context of the backend work is 50,000 tokens. The log entry is 50 tokens. The frontend agent gets exactly the information it needs at 0.1% of the context cost.
When to Read the Log
Agents should read the shared log at natural checkpoints - before starting a new subtask, after completing a subtask, and when they encounter something that depends on another agent's work. Not continuously. Continuous log monitoring wastes context.
Fazm is an open source macOS AI agent. Open source on GitHub.