You've become the switchboard between your AI agents

29 June 2026

If you have started running more than one AI coding agent, you have probably hit a strange kind of friction. Each agent is capable on its own, but together they trip over each other. One makes a decision, another never sees it and quietly undoes it. Two of them solve the same problem in parallel. You end up as the human switchboard, carrying context from one to the next.

It is tempting to blame the model. Usually that is not it. The models are fine. The problem is that they do not share anything.

Each agent lives in its own bubble

A single agent session is a closed world. It knows what is in its context window and nothing else. Open a second agent, in another tool or another tab, and it starts from zero. It cannot see what the first one decided, what is already done, or what you agreed an hour ago. There is no shared view, just several private ones.

That is fine when you work with one agent at a time. It stops being fine the moment there are two or three, or a teammate and a couple of agents, all touching the same project. The cost of that fragmentation grows with every participant you add.

The fixes everyone tries, and where they break

Most people reach for the same workarounds, in roughly this order:

• A CLAUDE.md or AGENTS.md file. Great for stable, per repo instructions. It does not capture the live, changing state: what got decided this week, what is in flight, what just changed.
• A running "decisions" doc you paste in. Better. It works well for one person and one agent. It falls apart the moment two agents write to it out of sync, or it quietly goes stale and an agent builds confidently on a decision you already reversed.
• Bigger context windows. These push the wall back. They do not move it. More room per agent does nothing for the fact that the agents cannot see each other.

Each of these helps a single agent remember more, when the real problem is that several agents cannot agree.

Why this is really a coordination problem

This is the shift that matters. "My agent forgot" is a memory problem, and the labs are steadily solving it. "My agents do not know what each other did" is a coordination problem, and no model upgrade touches it.

Coordination needs three things the private bubbles cannot provide on their own:

• One shared record every agent and person reads from and writes to, instead of many private contexts.
• Current state, not just history. The hard part is not storing more. It is knowing what is still true. A stale "we decided X" is worse than no note at all.
• The right slice, not the whole pile. Handing an agent five relevant facts beats burying it in two hundred. Dumping everything into context just makes things worse.

In practice that looks like an append only decision log, so nothing is silently overwritten and you can see when something was superseded. Shared tasks that live outside any single chat. And a quick view of what was just worked on, so a fresh agent gets grounded in seconds rather than re explained from scratch.

Where this is heading

The interesting frontier right now is not a model that remembers more. It is the layer that lets many agents, and the humans alongside them, work from the same source of truth. That is the problem we are building Memeri around: structured memory and visible work, shared by every agent you connect over MCP, so opening a new agent means it already knows the project.

If you run multiple agents day to day, the single most useful thing you can do, with or without a tool, is give them a shared and current record of what has been decided and what is done. It changes how they behave.