How specialists coordinate their work.
Eighteen team shapes. Thirty-one supporting practices. Choose how work flows, how it is challenged and when it returns to you.
Overview
Fan-out, fan-in
Several specialists work the same case in parallel, the Director merges.
Independent contributions converge into one answer. Select a role to trace its connections.
The shape is only the beginning.
Combine a topology with practices for review, memory, cost control and your decisions. These are design patterns, not claims that every pattern is available in every OS.
31 of 31 supporting practices
Diff approval
The gate shows what changes, not a paragraph, and the approval is bound to that exact diff.
Overnight precompute
The expensive part runs off-hours so the interactive path stays fast and cheap.
Batched approvals
Gates are collected and shown as one review at a set time instead of interrupting all day.
Dry run then apply
The OS produces the exact change it would make, a person approves it, then it is applied.
Human pairing
The person works in their own tool, the OS follows along and prepares the next artifact.
Chain of verification
The answer becomes checkable claims, each claim is verified alone, then the answer is rewritten.
Skill selection
The OS loads only the skills the job needs, judged per turn.
Two-person rule
Two different people, or a person and an independent check, approve before an effect lands.
Escalation ladder
One attempt per level, with a defined promotion when it fails or times out.
Rubric grading
Output is scored against a named rubric; anything under the bar goes back with the failed criteria.
Long-running timers
Steps wait for a date, an SLA, or an outside answer, and wake on their own.
Reflexion
After a failure the agent writes the lesson to memory and retries with it in context.
Checklist run
A fixed list of checks runs every time, with evidence attached to each line.
Circuit breaker
After repeated failures the branch opens and work falls back to a safe path.
Batch window
Items are collected for a period and handled as one batch, to cut cost and noise.
Memoized subgraph
A deterministic part of the graph is cached by input hash and skipped when nothing changed.
Best of n
The same task is sampled several times and a judge with a rubric picks the winner.
Red team and blue team
One team tries to break the output, the other fixes it, until the attacks run dry.
Contract net
The Director publishes a task, agents bid with cost and confidence, the best bid wins it.
Anomaly escalation
A detector flags the unusual, a specialist investigates, a person decides.
Interview loop
The OS asks a short list of questions, one at a time, until it can start.
Sourced writing
Retrieval gathers passages, the writer may only use those, and cites them.
Curriculum
The OS solves an easier version first, then uses that to attack the real one.
Quorum join
Several agents answer the same question, the run continues when k of n agree.
Dead letter queue
Work that failed every retry is parked with its context for a person to inspect.
Experiment runner
Variants are assigned by rule, results are collected, and a person picks the winner.
Provider fallback
Providers are tried in order until one answers within budget.
Shadow mode
The OS does the whole job but nothing is applied, and a person compares it with what really happened.
Self consistency
One question is answered several times independently and the majority answer wins.
Rate-limited queue
Work is buffered and drained at a rate the connected tools can take.
Structured debate
Two agents argue opposite positions under rules, a chair decides.