12 -Swarm gallery

Product decision

A product manager frames the user problem and shapes scope; a principal engineer grounds it in architecture and delivery risk. An orchestrator resolves their views into one call - Proceed, Defer, or Reject - grounded in both user value and engineering feasibility.

The roster

Two agents run in parallel. Each contributes a structured stance; the orchestrator synthesizes them.

Agent	Role
`product-manager`	Frames user value, controls scope, sequences delivery, and names the clearest next product decision.
`principal-engineer`	Grounds decisions in architecture, implementation risk, operational concerns, and realistic delivery feasibility.

Goal & decision

The preset's goal is: Reach a decision-ready recommendation grounded in user value and engineering feasibility.

The decision vocabulary is Proceed / Defer / Reject. Every agent stance maps to one of these words; the synthesis surfaces the majority call (or the orchestrator's resolution when the two agents diverge).

Resolution is set to orchestrator, so after each round the bundled orchestrator agent reads both structured outputs and produces the directive for the next round - or, on the final round, folds them into the synthesis.

Run it

The first positional argument is the number of rounds (1-3). Pass your question as the second argument, then name the preset:

agent-swarm run 1 "Should we adopt server components?" \
  --preset product-decision

For a two-round run that lets the agents refine their positions:

agent-swarm run 2 "Should we adopt server components?" \
  --preset product-decision

Note

Real harnesses can take longer than the default 120 s dispatch timeout. Bump --timeout-ms 300000 for deeper runs, or add it to .agent-swarm/config.yml. Use --quiet for one-line-per-event output in CI.

What the synthesis looks like

When the run finishes you'll find a synthesis.md in .agent-swarm/runs/<timestamp>-<slug>/. The verdict word comes straight from the preset's decision vocabulary; confidence is reported as a level, not a fabricated percentage.

Synthesisproduct-decision · round 1 · 2 agents

Proceed - server components fit the migration if routing is staged incrementally and data-fetching boundaries are settled first.

consensus2 of 2 agree

confidencemedium

Still openDoes the data-fetching layer change before or after the routing migration?

Confidence is low, medium, or high. The "Still open" row surfaces deferred questions that neither agent resolved - a signal of where more information is needed before acting.

Related presets

Three presets share the same goal and decision vocabulary as product-decision:

product-triad: Expands the roster to three agents - product-manager, product-engineer, and product-designer - adding a user-experience perspective alongside product and engineering. Same Proceed / Defer / Reject vocabulary, same orchestrator resolution.
product-decision-codex: The same two-agent product decision swarm running on the Codex harness. Identical goal and decision vocabulary; swap in when your team uses Codex instead of Claude.
product-decision-opencode: The same two-agent product decision swarm running on the OpenCode harness. Identical goal and decision vocabulary; swap in when your team uses OpenCode.

Tip

To mix harnesses or add a third voice, define your own preset in .agent-swarm/presets/. See Composing a swarm for the full authoring guide.

Where next

Overview - the full gallery of bundled swarms.
Quickstart - run product-decision end to end in five minutes.
Composing a swarm - write your own agents, presets, and decision vocabulary.
CLI reference - every command, flag, and run artifact.