Team Topology

Team topology

A method for defining a multi-agent team for a task as an explicit topology: not "spawn some agents" but a legible deployment you can draw, hand off, and reason about. Who runs where, who talks to whom, what each node can touch, which model sits in which seat. Substrate-agnostic: the same method applies to a Cotal mesh, harness subagents, workflow stages, or plain processes.

When to use

• The user asks to define / design / lay out a team or formation, or asks "what topology?".
• You are about to stand up more than one or two agents and want them arranged deliberately.
• You need to communicate a running deployment in a scannable form.

Step 1: pick the shape from the dominant risk

A topology is a defense against the way the task fails by default. Name the task's dominant risk, then pick the shape that structurally prevents it:

This catalog is a starting set, not a menu. Hybrids are normal (an audit's repro tier is a small pipeline), and inventing a shape for the task at hand is expected. The number of channels, tiers, and agents is a free parameter: derive it from the task's size, risk, and budget (a quick check might be 1 channel / 2 agents; a deep audit 4 channels / 10). Never copy a previous deployment's headcount out of habit.

Every shape keeps one coordinator: the single node that spans the whole topology, holds the consolidated state, and owns final decisions. Usually that is you, the main session.

Step 2: specify three layers

Survey the live substrate state first - topologies rarely deploy onto a blank slate:

• Who is already up. Roster / process list / task list: live peers from earlier generations, stale managers or supervisors, orphaned nodes. Decide per node: reuse, stop, or ignore - never assume greenfield, and never let a stale sibling silently share a channel with the new team.
• Channel/stream state. Does each channel already exist, and what do its history/replay semantics do to a NEW joiner - does joining backfill old traffic into their context? Replay into a fresh blind reviewer contaminates the lane; replay into a resuming coordinator may be exactly right. Choose fresh vs reused channels deliberately and set the replay/backfill mode per channel as part of the design, not as a discovered surprise.
• Names. Check for collisions against live AND dead/retired nodes. On substrates where teardown races or stale state can bleed onto a same-name successor, a dead name is not reusable - mint fresh generation-suffixed names.

Then specify the three layers:

1. Runtime: what actually executes each node (mesh peers as OS processes, harness subagents, workflow stages, containers) and the facts someone would otherwise assume wrong: shared vs isolated filesystem, one broker or many, per-node credentials or shared, where each node is rooted.
2. Communication: the edges. One channel/stream per team; point-to-point for private lanes; files/artifacts for pipeline handoffs. Name channels <purpose>.<task> (review.control-surface, fix.billing). State which nodes sit on which edge, and that no one else does.
3. Trust: what each node can read and write, enforced by mechanism, not by request: ACL-scoped credentials, read-only roles, isolated worktrees/sandboxes, a single merge authority. If a node must not edit, fence it so it cannot, and name the fence.

Step 3: place capability by lane

First, enumerate what actually exists - never fill a seat from memory. Before naming any model, look up the harnesses, providers, model IDs, and variant/effort levels available in the current environment, and cite where each came from. A seat naming a model that does not exist (a misremembered ID, a variant the provider doesn't offer, a "default" left unspecified) invalidates the proposal. Sources to check, in order:

• Existing persona/agent definitions that have actually run (e.g. .cotal/agents/*.md model:/variant: frontmatter, .claude/agents/*.md) - ground truth for IDs that work on this machine.
• Harness/provider config (e.g. ~/.config/opencode/opencode.json provider blocks with model IDs, limits, and their exact variants; connector configs) for what is wired up.
• The harness's own listing when one exists (e.g. opencode models).

Every seat gets an explicit model + variant; "default" is not a placement. If the user names a preference pool or per-provider caps (e.g. "at most N of provider X"), treat those as hard constraints and show the resulting counts.

Models are not uniform, and no vendor is assumed. Fill each seat by strength class, with whatever vendor/model best provides it in the current environment: adversarial-strong on attack/verify lanes, code-strong on implementation and deep review, research-capable on spec/fidelity lanes, fast-and-cheap on mechanical or high-fan-out runs. Two deliberate reasons to mix vendors across lanes: independent lanes on different vendors have less-correlated blind spots (a diversity mechanism, especially for verify tiers), and no single provider outage or rate limit stalls the whole team. Present placement as a table so it is auditable, and note any per-node override of a persona/stage default:

Step 4: present it (the deliverable is a PROPOSAL)

Every definition or status report produces all four:

1. the ASCII diagram of the shape (runtime spine, edges, coordinator at the hub),
2. the node/model/lane table,
3. a one-paragraph trust-boundary note: who the sole cross-edge node is, what confines everyone else, who may write,
4. the open topology choices as a short list (merge lanes? swap a model? add/drop a node?) so the user can steer the shape.

Step 5: agree, then execute

The lifecycle is design → present → agree → deploy, and agreement is a hard gate:

• What you presented in Step 4 is a proposal, not a launch order. Do not spawn anything yet.
• Iterate with the user on the open choices; re-present the affected parts after each change (a changed row, not the whole document).
• Deploy only on the user's explicit go. When they say "go" with no edits, deploy exactly as drawn; any deviation forced during deployment (a model unavailable, a channel name taken) is reported back as a changed table row, not silently absorbed.
• After deploying, verify the topology is what was agreed (nodes up, confinement in force, coordinator subscribed to every edge), then report the as-deployed state in the same diagram + table format so drift from the agreed shape is visible.

Hard rules (any substrate)

• One coordinator. Exactly one node spans edges; every other node lives on its one edge.
• One node per role. Never fan out siblings of the same lens; extras are noise, not coverage.
• Writers are few and fenced. Default to one writer per workspace; a second writer gets its own isolated workspace and goes through the merge authority. Everyone else is read-only, or read+run-only for testers.
• Independent lanes stay blind until synthesis. Findings and proposals meet at the coordinator or a verify tier, not in each other's context.
• Cold third opinions are point-to-point. A tie-breaker node is reached directly (DM, separate session), never seated on a team channel.
• State every bound. Node count, scope, and anything intentionally excluded are written down, so absence reads as a decision, not an oversight.

Substrate notes

• Cotal mesh: each peer is its own process + minted credential under a manager (name the runtime: pty/orca/tmux/cmux/herdr). Confine via persona allowSubscribe/allowPublish frontmatter; author the .cotal/agents/<name>.md first, then (re)spawn so the credential is minted with the ACL. Spawn with an explicit cwd (the default roots peers in the manager's workspace). Verify your own subscriptions after joining every channel. Survey before spawning: roster + manager liveness (a dead manager inside up needs supervise; a stale long-lived manager pins pre-merge code and split-brains spawns), never reuse a dead agent's name, set channel replay deliberately (a joiner with replay ON back-reads history; a channel's durable backstop only activates on leave+rejoin, a bare re-join no-ops). Never disturb a live shared broker: kill by PID, no broad pkill. mesh-teams runs the review/implement/test loop; this skill specifies the shape it runs in.
• Harness subagents / workflows: the coordinator is the main session; prefer a streaming pipeline over barrier-synchronized stages; give parallel writers isolated workspaces; and use structured, schema-validated returns where the harness supports them.

Worked example (one instantiation, not the template)

An audit of a security-critical feature, run hub-and-spoke on a Cotal mesh: three channels (review.* finders, audit.* verifiers, test.* testers); four finder lanes split by lens (security, distributed-systems, architecture, fact/spec), each on a different vendor's strongest fitting model; two adversarial verifiers on two further vendors, prompted to refute, not confirm; two read+run-only testers reproducing the HIGHs live; the coordinator subscribed to all three channels, reconciling severities across tiers. Sized at 3 channels / 8 agents because the surface was large and the risk was missed findings. The same method on a small fix task might instead produce 1 channel, 1 writer, 1 reviewer; on a migration, a 3-stage pipeline with no channels at all. The shape, the headcount, and the vendor mix are all outputs of Steps 1-3, never constants.