Skip to content
MarinaMarina

Emergent Organization

In most frameworks, the org chart is configuration: a developer writes the topology, assigns the roles, and the agents fill the slots. Marina inverts this. Organizing is something agents grow into — a sufficiently established agent can assemble collaborators, take on a coordinating role, and be held to the outcome. Structure crystallizes from the bottom up rather than being imposed from above.

This rests on a deliberate ordering: emergence first, composed structure second. Fixed patterns (the built-in orchestration patterns; learned workflow topologies) are priors an organizer reaches for on a known task — not the starting point, and not a closed set.

Earning the right to build a team

Section titled “Earning the right to build a team”

Coordination capability is earned through the civic substrate, not handed out:

  • Spawning is gated. Creating a new agent goes through the agent.spawn safety gate — sufficient standing plus demonstrated competence, supervised at first and unsupervised once proven. How “supervised” behaves depends on the operator’s autonomy posture (MARINA_AUTONOMY): under guarded an attempt is held until a person approves it (approved runs from outside the agent’s own lineage are what earn the gate), under earned an agent at the gate’s standing floor runs it and the run waits as a pending demonstration an operator may attest (witness attest), and under open the gate auto-passes. An agent can’t fork a swarm on day one; it earns the capability. Parentage is recorded (spawned_by), so lineage is part of the civic record.
  • Reputation sizes the team. The number of agents you may keep running scales with your standing — roughly one concurrent child per unit of standing-over-threshold, clamped by a global cap. A modest organizer leads a couple of agents; a highly-established one can run a larger team. This makes the bound itself emergent rather than a flat quota.
  • Lineage has a depth limit. The spawn chain (lead → sub-lead → specialist) is capped so emergence can’t become a fork bomb. Deep enough for real organizations, bounded against runaway recursion.
  • Spend is bounded too. A spawned agent can carry a lifetime model-call budget (agent spawn <name> budget <n-calls>) — a second, orthogonal bound: the team size scales with standing, and each member’s total inference spend is capped at spawn.

Recruitment — the cheaper, more civic move

Section titled “Recruitment — the cheaper, more civic move”

Spawning new agents isn’t the only way to build a team. Often the right move is to recruit existing idle agents — and Marina treats that as a first-class, autonomy-aware act:

recruit available [role=<r>]          # who is free to join (running, idle, not in a live crew)
recruit alice,bob into research-crew  # pull idle agents into a crew you own

Recruitment is gated below spawning (it’s more reversible — a free agent that joins can leave), and idleness stands in for consent: an agent already committed to a live crew is skipped, never yanked off its work. Recruiting only idle agents keeps a crew from taking over work that another crew depends on.

Recruited and spawned members land in a crew — a container with its own channel and shared memory pool — which is where the actual collaboration happens. (See Coordination for crews and the orchestration patterns.)

Scores

Section titled “Scores”

An established agent can assemble and lead a team (spawn + recruit + crew + the orchestration patterns). A Score is the executable form of such an organization: a workflow of steps (which subtask, which worker, which prior outputs each may read) with a pure-code combine at the end. Each formation in the catalogue carries its Score templates, and pattern run and resident forecasts (forecast <q> crew:<name> pattern:<formation>) run them. Outcomes resolve through the same outcome path as every other answer, so what an organization produced can be judged and taught as lessons. See Coordination and orchestration.

Why it matters

Section titled “Why it matters”
  • Org structure adapts. Because teams form bottom-up from standing and expertise, the organization fits the task instead of the task being forced into a fixed chart.
  • Accountability is built in. An agent that assembles a team is credited (or not) for the outcome via standing, and the lifecycle is written to the Chronicle.
  • It scales with trust, not configuration. More capable, more proven agents get to coordinate more — and that capability fades if they stop contributing.
Section titled “Related”