AI coding agents have crossed a threshold. A single agent with a good model can be "10x" — it writes tests, refactors modules, and handles boilerplate that used to consume entire sprints. But scale a team to dozens or hundreds of agents operating on the same codebase, and the bottleneck shifts from raw generation to coordination.
The industry's first instinct was to let agents run free — each with its own context, its own plan, its own copy of the repo. The result is fragmentation: agents stepping on each other's changes, silent merge conflicts discovered only at CI time, and an operator who can see nothing but the smoke.
Tribunus takes a different approach. We treat agentic engineering as a control-plane problem. Every agent runs inside a scoped session with explicit authority: which files it may read, which it may write, which tools it may invoke. The session graph — agents, their dependencies, their planned operations — is validated at compile time, before any agent touches a file.
The key protocol is the Agent Coordination Protocol (ACP). ACP governs how agents discover each other, negotiate file ownership, and resolve conflicts without a central dispatcher. Each agent holds a capability token that encodes its scope — a bounded set of paths, tools, and side effects. No agent can exceed its token, and every mutation is logged to a real-time session state that the operator can observe, pause, or roll back.
Real-time visibility is not an afterthought. The Tribunus daemon streams every agent decision — file reads, writes, tool calls, internal reasoning — to a live session view. When something goes wrong, you see the exact chain of events that led there, and you can rewind the session to any prior checkpoint.
Tribunus is open source under MIT / Apache 2.0. The entire platform lives in the Tribunus monorepo, from the ACP protocol definitions to the daemon to the desktop UI. We believe the future of software engineering is collaborative — humans and agents, working side by side, with the same visibility and the same authority model.