HomeBlog › The CEO agent does not talk to fifty workers
Blog

The CEO agent does not talk to fifty workers

Aug 17, 20264 min readBy the Rysh team

Graph Engineering pattern 10 of 15: the Hierarchical Fleet Graph. Hierarchy as coordination-cost compression, context locality — and a measured 22-fleet run against rysh's own roadmap.

The CEO agent does not talk to fifty workers. Somewhere between five agents and fifty, every flat topology drowns: the manager's context fills with status instead of decisions, and the star (pattern 1) collapses under its own fan-out. The fix is the one human organizations found centuries ago — hierarchy. Not as bureaucracy: as coordination-cost compression.

Context for new readers: pattern 10 of 15 in the Graph Engineering series. rysh is an agentic terminal multiplexer; this is the pattern we run on rysh itself, so this article carries measured numbers rather than projections.

The topology

                CEO
                 |
          Fleet Coordinator
         /       |        \
       FM1      FM2       FMn
        |        |         |
     Fleet1   Fleet2    Fleetn
     /|..\    /|..\     /|..\
    W W   W  W W   W   W W   W

A CEO agent owns the goal and its top-level decomposition. A fleet coordinator allocates workstreams to fleet managers and handles cross-fleet dependencies. Each fleet manager runs its own bounded fleet — internally just a star or a fan-out/fan-in — and reports upward in compressed form: decisions needed, blockers, deltas against plan. Never raw activity.

Two properties, one shape

Compression. Information is summarized on the way up and refined into instructions on the way down, exactly as in a human organization and for exactly the same reason: coordination cost grows super-linearly with direct relationships. A manager that forwards raw worker output upward is overhead without function.

Context locality. Each fleet's workers share deep context about their workstream; the fleet manager owns the boundary; the layers above deal only in interfaces between workstreams. This mirrors the codebase structure the fleets are usually building — and it bounds the blast radius of any single agent's confusion to its own fleet.

The measured run

On 2026-08-11, one command registered 22 fleets against rysh's own roadmap and brought three up — four agents each, twelve live Claude sessions, twelve worktrees, a tab and a board per fleet, zero failures. The other nineteen were refused by a concurrency cap the registry itself enforces. Two details make that a claim rather than a boast: registered is free — the cap counts only running fleets, so a driver may register everything it discovers and promote a bounded few; and over-cap is refused, not queued, with the refusal naming the fleets already up — so a typo and a full cap do not read alike.

Whose chart it is — honestly

The panes, the given-names, the persistent pane metadata, the worktree isolation and the pane-to-pane addressing all ship in the rysh binary. The ceo→manager→worker chart built on top of them is our in-house driver tooling, and we name it as such: the binary's fleet registry knows a fleet's name, state, board and members — it does not know that one member manages another. The topology is engineered by the organization on primitives the platform guarantees. That is Graph Engineering practiced on ourselves.

Where it breaks

The dimension answers

Dimension Answer
Nodes CEO, coordinator, fleet managers, worker fleets
Edges Vertical delegation/reporting; cross-fleet edges at the coordinator
Cycles Plan-revision loops between adjacent layers
Humans Goal-setting above the CEO agent; approval on external effects
Exit rule Per-fleet exit rules roll up; CEO closes the epic against its goal

Honest scope

Next: pattern 11, the Research–Synthesis Graph — broad exploration, central synthesis, and an experimenter that keeps conclusions honest.


Building this with design partners → rysh.ai/design-partner

Try Rysh

Every pane is a shell and an AI agent — install takes one command.

Get started free →