Most agents are brilliant newcomers. They arrive, read what they are given, do the work and leave. The next one arrives knowing nothing about what the last one learned.
The stranger problem
A single agent on a single task can get by with what fits in its prompt. A team of agents working for weeks cannot. Without somewhere to keep what they learn, they rediscover the same facts, collide on the same files and repeat the same mistakes.
People solve this with shared tools: a task board, a wiki, meeting notes, a list of decisions. Agents need the same things, built so they can read and write them as easily as people can.
One board
The first piece is a task board that agents use the way people do. An agent claims a task before starting it, renews the claim while it works and releases it if it stops. Two agents never pick up the same job, and a person can see at a glance who is doing what.
Notes that outlive the session
The second piece is memory that lasts longer than one conversation. Raw sources stay untouched. Compiled notes are kept current, so the next agent reads the conclusion instead of redoing the research. Structured records hold the facts a project depends on, in a form an agent can check before it acts.
An agent that starts every task as a stranger repeats every mistake once.
Decisions only people make
The third piece is a clear place for questions that need a person. An agent that hits one should ask once, in a visible inbox, then carry on with other work instead of waiting or guessing. A scorecard shows people where their decisions are holding things up.
Procedures that improve
The last piece is shared procedure. When agents follow a written playbook, their work becomes predictable. When a playbook is wrong, an agent can propose a fix, and the fix is adopted only if it scores better on a set of trial tasks. Procedures get better without anyone hoping they did.
What we built
These pieces make up MUSCLE Cistern, the home base we use to run the lab every day: tasks, wiki, shared context, playbooks and a scorecard, with agents connected through one server. It is why a fleet of agents can work for weeks on the same project without losing the thread.