Strategy Reference
Auto-generated from the strategy-manifest registry (
src/orchestration/strategy-manifest-registry.ts, the single source of truth the router reads). Do not edit by hand — runpnpm docs:strategiesto regenerate.
run is the canonical entry point: give it a goal and the
MetaOrchestrator routes to one of these strategies for you. You rarely need to
pick a strategy by hand. The specialized tools below remain available as
force-strategy escape hatches — pass forceStrategy to run, or call the
strategy’s entrypoint tool directly — for when you already know exactly which
engine the work needs. Each row tells you when forcing is the right call.
nexus-agents has 8 routable execution strategies.
| Strategy | Entrypoint tool | When to force | Maturity | Authority | Executor |
|---|---|---|---|---|---|
consensus |
consensus_vote |
Force when a decision needs N independent voters rather than one model. | stable | advisory | wired |
dev-pipeline |
run_dev_pipeline |
Force when the goal is a code change that must pass the dev quality gate before it counts as done. | stable | suggest | wired |
graph-workflow |
run_graph_workflow |
Force when the work is an explicit dependency graph with conditional edges (a predefined workflow template). | beta | suggest | fail-closed |
orchestrate |
orchestrate |
Force when the work needs multi-agent orchestration patterns rather than a single linear pipeline. | beta | suggest | fail-closed |
pipeline |
run_pipeline |
Force when the work fits a templated multi-stage pipeline rather than a single model call. | stable | suggest | wired |
research |
run_pipeline |
Force when the goal is research-led (gather, synthesize, compare) rather than a code change or decision. | stable | suggest | wired |
single-shot |
delegate_to_model |
Force when the goal is a one-shot ask that needs no pipeline, gate, or multi-step plan. | stable | suggest | fail-closed |
spec |
execute_spec |
Force when building a greenfield project from a written spec, not a plain goal string. | beta | suggest | fail-closed |
Per-strategy detail
consensus
Multi-perspective decision routed to a consensus vote.
- Entrypoint tool:
consensus_vote - When to force: Force when a decision needs N independent voters rather than one model.
- Maturity tier: stable
- Authority tier: advisory
- Executor: wired (runs inline with
execute:true)
dev-pipeline
Code change run through the dev gate (test / lint / typecheck).
- Entrypoint tool:
run_dev_pipeline - When to force: Force when the goal is a code change that must pass the dev quality gate before it counts as done.
- Maturity tier: stable
- Authority tier: suggest
- Executor: wired (runs inline with
execute:true)
graph-workflow
DAG / conditional-edge workflow execution.
- Entrypoint tool:
run_graph_workflow - When to force: Force when the work is an explicit dependency graph with conditional edges (a predefined workflow template).
- Maturity tier: beta
- Authority tier: suggest
- Executor: fail-closed (no inline executor wired yet; routing never auto-selects it)
orchestrate
Pattern-based multi-agent orchestration (wave / aflow / puppeteer).
- Entrypoint tool:
orchestrate - When to force: Force when the work needs multi-agent orchestration patterns rather than a single linear pipeline.
- Maturity tier: beta
- Authority tier: suggest
- Executor: fail-closed (no inline executor wired yet; routing never auto-selects it)
pipeline
Multi-stage templated work (audit / general) via the pipeline engine.
- Entrypoint tool:
run_pipeline - When to force: Force when the work fits a templated multi-stage pipeline rather than a single model call.
- Maturity tier: stable
- Authority tier: suggest
- Executor: wired (runs inline with
execute:true)
research
Research-heavy work routed through the research pipeline.
- Entrypoint tool:
run_pipeline - When to force: Force when the goal is research-led (gather, synthesize, compare) rather than a code change or decision.
- Maturity tier: stable
- Authority tier: suggest
- Executor: wired (runs inline with
execute:true)
single-shot
Trivial single-step task delegated to one model.
- Entrypoint tool:
delegate_to_model - When to force: Force when the goal is a one-shot ask that needs no pipeline, gate, or multi-step plan.
- Maturity tier: stable
- Authority tier: suggest
- Executor: fail-closed (no inline executor wired yet; routing never auto-selects it)
spec
Greenfield project built from a markdown spec document.
- Entrypoint tool:
execute_spec - When to force: Force when building a greenfield project from a written spec, not a plain goal string.
- Maturity tier: beta
- Authority tier: suggest
- Executor: fail-closed (no inline executor wired yet; routing never auto-selects it)