Engineering
The anatomy of an agent team: Orchestrator, Workers, and Critic, explained
"Multi-agent orchestration" gets used to describe almost anything with more than one model call in it. On Claws, it means one specific, enforced structure, the same one whether you're running a five-post content calendar or a B2B sales pipeline. This is the full architecture, worked through with a real template so none of it stays abstract.
5 min
read
·
The shape, exactly
Every Claw, official or custom, is 1 Orchestrator + 2 to 6 Workers + 1 Critic, 4 to 8 agents total. This isn't a default you can turn off. The visual builder won't let you deploy a team missing either the Orchestrator or the Critic; both are required on every single team, no exceptions.
The three roles
The Orchestrator is the only agent that talks to you. It reads what's due, delegates tasks to the right Workers, collects their output, and routes it to the Critic before anything reaches you. It's the only agent that binds to your messaging channels. It runs on a fast, cheap model, since routing doesn't need deep judgment.
Workers (2 to 6) are the specialists, research, writing, analysis, outreach, whatever the Claw actually does. Each one has its own identity and its own job. Most run on a mid-tier model; creative roles often run on the strongest one instead. You never see a Worker directly. You only see what survives review.
The Critic (exactly one) reviews every Worker output before it ships. It runs on the strongest available model, since judgment matters more here than anywhere else in the team.
How they actually talk to each other
Agents don't just message each other freely. The Orchestrator spawns Workers, and the Critic, as subagents through a specific tool call. Results announce back automatically once each subagent finishes its task. That's the entire communication model: delegate, work, announce back, review, decide.
One cycle, start to finish
On every scheduled trigger, the Orchestrator runs the same sequence:
Checks the task list for anything due right now.
Decomposes the work and spawns the right Workers, each with a 5-minute timeout.
Collects every output and sends the batch to the Critic.
Processes the verdict. Approved work gets compiled and presented to you. Rejected work goes back to the Worker with specific feedback.
Logs the whole cycle to that day's memory file, whether it succeeded or not.
If a piece of work fails review three times in a row, the Orchestrator stops looping and presents the best attempt anyway, flagged clearly as below threshold, instead of retrying forever or quietly lowering the bar.
What the Critic actually scores
The default review runs against four criteria, weighted:
Accuracy (30%): facts, numbers, and names correct, no hallucinations.
Completeness (25%): the task is fully addressed, no gaps.
Quality (25%): professional, well-structured, matches your tone.
Risk (20%): nothing that could embarrass you or create a legal, compliance, or reputational problem.
A 6 out of 10 is a reject. Only 8 and above passes. Individual Claws can layer domain-specific checks on top, a bias check for recruiting, a Fair Housing check for real estate, but these four axes are universal across every team.
Memory across three layers
Session memory is ephemeral: working context for the task in front of the agent right now, gone once the session ends.
Daily memory is a dated file, retained 30 days before it's archived. The Orchestrator reads today's and yesterday's at the start of every session, so it doesn't start cold.
Long-term memory never gets deleted. Your business context, recurring patterns, and "do not do this" preferences live here, and every agent reads it before doing anything.
A shared pipeline file also tracks the team's current workflow state, so Workers picking up a task mid-process aren't starting from nothing.
Worked example: a B2B Sales Machine
B2B Sales Machine ships as 7 agents. Its documented pipeline runs ICP research, enrichment, five-email sequences, 7-category reply classification, and pipeline management, five distinct stages, which maps naturally onto one Orchestrator, five Workers, one Critic:
ICP research: builds the ideal-customer profile to target.
Enrichment: turns a raw lead into usable contact and firmographic data.
Sequence writing: drafts the five-email outreach sequence.
Reply classification (7 categories): sorts responses into the right one.
Pipeline management: tracks what's actually sitting in the funnel.
On a heartbeat, the Orchestrator delegates a batch of enrichment tasks, waits for the outputs, and sends them to the Critic. The Critic checks the enriched data against the Accuracy axis, checks the sequence drafts for CAN-SPAM compliance under Risk, and rejects anything that doesn't clear an 8. Only what passes reaches you, as one compiled update, with a hot-lead alert flagged separately if a reply classifies as ready to buy.
Swap the domain and the Workers change entirely. The shape underneath, one router, a bench of specialists, one reviewer who never wrote what it's reviewing, doesn't.
