Skip to content
Marina Marina

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 ten 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 needs a witness-granted window, under earned it runs freely pending attestation, 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 birth.

Recruitment — the cheaper, more civic move

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

Spawning new minds 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. That autonomy guard is the difference between building a civilization and commandeering one.

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 ten orchestration patterns.)

Where this is going: the Score

Section titled “Where this is going: the Score”

Today an established agent can assemble and lead a team (spawn + recruit + crew + the orchestration patterns). The frontier — inspired by recent work on learning to orchestrate agents in natural language — is the Score: a dynamic, executable, scored workflow an organizer composes on the fly (which subtasks, which worker, which prior outputs each may read), with topology (chain, tree, best-of-N) emerging from results rather than being pre-declared. The non-gradient learning loop that would score these organizations — pairing an attempt with its resolved outcome into a recallable note — already exists in the substrate. Turning emergent organizations into inheritable, mutable priors is the direction; the earned-assembly foundation it builds on is real today.

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”