Building an AI brain for a team of engineers, not just for yourself

The system that makes one engineer extraordinary does not make a team of engineers extraordinary. It usually makes them inconsistent. A personal brain can be opinionated, messy and implicit, because one person arbitrates everything. The moment five engineers and a fleet of agents share a memory, three new problems appear: who decides what is true, how agents avoid trampling each other, and how knowledge stays current instead of quietly forking. A team brain is the machine that answers those three questions.

Context engineering is the delicate art and science of filling the context window with just the right information for the next step.

— Andrej Karpathy, June 2025

That principle holds for one person and for a hundred. What changes with a team is who curates. On your own, you are the editor. In a team, editing has to become a process, or the shared context fills with five people's half-true notes and every agent inherits the confusion at machine speed.

The single most important rule. Anyone, human or agent, can read the shared brain at any time: conventions, decisions, playbooks, the wiki. But nothing enters it without review. Memory writes are pull requests. A correction one engineer teaches their assistant becomes a proposal, not an instant global fact, because a wrong fact in a shared brain is a defect that replicates into everyone's work, including the agents'.

A wrong fact in a shared brain is a defect that replicates into everyone's work.

The fear people have about multiple agents is the same fear they had about multiple engineers, and the answer is the same machinery, applied with more discipline. Each agent works on its own branch in its own worktree, an isolated copy of the repository. Nothing it does touches anyone else's work until it survives review and continuous integration. Maximum parallelism, zero collisions.

git worktree add ../agent-payments -b agent/payments-retry
git worktree add ../agent-reporting -b agent/reporting-export

A team brain is versioned like code, because it governs code. The conventions file, CLAUDE.md or the vendor-neutral AGENTS.md, is checked in at the root, so every engineer's assistant reads the same rules: how we name things, what we never do, where the bodies are buried. Shared capability ships the same way. A project-scoped MCP configuration means everyone's assistant gets the same tools with the same permissions, by default, on clone.

claude mcp add --scope project postgres -- npx -y @modelcontextprotocol/server-postgres
git add .mcp.json && git commit -m "Team-wide read access to the analytics db"

The highest-leverage thing a senior engineer can do in this model is no longer writing the code; it is writing the playbook that lets every agent do the work to their standard. A skill, a markdown procedure with the exact steps, checks and edge cases for a recurring task, turns one person's judgment into the whole team's default. Deploy runbooks, review checklists, incident triage: written once, executed by anyone's agent, improved by pull request like everything else.

One person's judgment becomes the whole team's default.

Code has tests; a shared brain needs the equivalent. A small evaluation set, real questions with known-good answers, runs whenever the brain changes: does the assistant still answer our top twenty questions correctly after this update? It is the difference between a knowledge base you trust and one you stopped trusting quietly in March. Tools like promptfoo make this an hour's setup, and it catches the failure mode no one watches for: knowledge that used to be right.

The load-bearing tools

Claude CodeAnthropicChecked-in CLAUDE.md conventions, project-scoped MCP, and subagents that work in isolated worktrees.Git worktreesBuilt into gitParallel working copies of one repository. The isolation primitive that keeps agents out of each other's way.GitHub CODEOWNERSGitHubRoute changes to the brain (conventions, skills, wiki) to the people accountable for them, automatically.Model Context ProtocolOpen standardShared tool access as configuration: one .mcp.json in the repo and every assistant has the same capabilities.PRpromptfooOpen sourceEvaluation harness for prompts and knowledge: assert the brain still answers correctly after every change.

The payoff is not that individuals get faster, although they do. It is that quality stops depending on who happened to do the work. Onboarding compresses from months to days because the institution's judgment is written down where both people and agents read it. The bus factor on tribal knowledge drops. And the organisation gains a new kind of asset on its balance sheet: an operating memory that improves with use instead of walking out the door at notice period.