Parallel research
Split broad repo, docs, or market investigation into independent read-only worker packets.
AGENTFANOUT
Provider-agnostic routing policy for safer multi-agent work.
AgentFanout is a portable orchestration skill and deterministic advisory router for teams building worker-capable AI runtimes. It helps decide when fanout is useful, how worker packets should be bounded, and what must stay under main-session control.
DIRECT_ANSWER
AgentFanout is a router and policy layer, not a standalone dispatcher.
AgentFanout classifies work, applies hard safety gates, shapes bounded worker packets, recommends provider routing, and defines validation expectations. The host runtime still launches workers. This keeps the project focused on portable orchestration policy rather than trying to replace every agent execution framework.
PUBLIC_BOUNDARY
What AgentFanout does and does not claim.
| Area | Supported Public Claim | Boundary |
|---|---|---|
| Routing | Classifies work and recommends whether fanout is worth the overhead. | It does not execute subprocesses or manage a queue. |
| Safety | Keeps private tools, secrets, git state, destructive actions, and final synthesis in the main session. | It does not make workers safe if a host runtime ignores the policy. |
| Providers | Provides provider-agnostic guidance for Codex, Claude, MiniMax, local LLMs, and future adapters. | It does not promise provider parity or live endpoint probing. |
| Validation | Requires validation criteria before delegated work is integrated. | It does not remove the need for tests, browser checks, source review, or human judgment. |
EXECUTION_MODEL
Four modes, one control principle.
AgentFanout can be used as a skill, a worker-first planning policy, a deterministic route script, or the foundation for a future host-integrated harness. In every mode, the main session remains the authority boundary and final integrator.
Hard safety gates
Secrets, private tools, destructive actions, git publication, and final synthesis are main-session-only.
Deterministic routing
The route script classifies task type, delegation value, provider fit, and validator requirements.
Validator-first integration
Worker output is useful only after the main session checks it against the real task and evidence.
USE_CASES
Built for controlled parallelism, not maximum automation.
Review and validation
Route independent checks to validators before the main session integrates findings.
Safety-aware coding
Separate bounded implementation work from private tools, git state, and final release decisions.
QUICKSTART
Install as a portable skill and inspect routing decisions.
AgentFanout is currently documented as v0.1.1. The public repository is MIT-licensed and includes the skill, routing script, tests, release notes, and safety model.
$ cp -R AgentFanout "$HOME/.pai/skills/AgentFanout"
$ python3 scripts/route.py "review twenty independent modules"
$ python3 tests/run_tests.py
FAQ
AgentFanout questions.
What is AgentFanout?
A portable skill and policy layer that helps host runtimes decide when multi-agent fanout is useful and how to scope it.
Does AgentFanout launch workers?
No. It is intentionally not a standalone dispatcher today. The host runtime launches workers.
What stays in the main session?
Private tools, credentials, git state, destructive operations, and final synthesis remain main-session-only.
How is it different from agent frameworks?
Frameworks execute workflows. AgentFanout supplies portable routing policy, hard gates, and worker packet shaping.
RELATED_TOPICS