Deliberation: Conditional Verdicts from Agent Debates

Where this fits

Imagine you are mid-sprint and two senior engineers on your agent team disagree: should the sync layer use Postgres or SQLite? You ask the agent to "battle this out" and get a definitive winner. That is not enough. The deliberation skill turns that debate into a structured Innovator → Devil's Advocate → Mediator pass, producing a conditional verdict that preserves the useful parts of both positions.

Why agents benefit

• Forces structured role separation: The agent does not drift into generic pros/cons; it must propose, challenge, and synthesize in distinct passes.
• Extracts salvageable fragments: Losing arguments are not discarded. The mediator identifies conditions under which the rejected option is still the right call.
• Avoids false binary outcomes: The final answer is a conditional verdict — "use Postgres unless offline write volume is critical" — instead of a winner.
• Routes depth-critical judgment deliberately: Optional persona jurors give domain-specific views without requiring an expensive model as the default substrate.
• Gives struggling agents a rope: Even when an agent cannot naturally challenge assumptions, the explicit devil's-advocate layer makes it do so.

Practical scenarios

Architecture choice between event-driven and polling

A developer needs to decide whether to replace a polling-based integration with WebSockets. The deliberation skill evaluates latency, operational complexity, and failure modes, then returns a conditional verdict: keep polling for low-frequency queries, migrate to streaming only when real-time updates cross a specific threshold.

Choosing between two skill proposals

Your agent has two competing approaches for a new reporting skill. Instead of merging feature lists, it runs Innovator and Devil's Advocate to expose hidden assumptions, then synthesizes a path that extracts the best of both.

API design tradeoffs

REST vs GraphQL for a public API. The devil's advocate highlights over-fetching and versioning issues; the mediator produces a conditional verdict that suggests GraphQL only for clients with complex query patterns, REST elsewhere.

Add it to your agent workflow

Install the skill with:

npx skills add chrono-meta/forge-harness --skill deliberation

Then trigger it with a natural prompt like:

Deliberation: weigh the pros and cons of Postgres vs SQLite for local-first sync, and give a conditional verdict.

The skill reads the session context, runs the three-layer structure, and returns a synthesized decision you can apply directly.

Compare before adopting

Before making it a permanent part of your harness, compare it against:

• Quality signals: Does the mediator output actually contain conditionals and edge cases, or is it a generic summary?
• Maintenance freshness: Check the GitHub repo's commit history and issues to see if the skill is actively maintained.
• Alternatives: A plain multi-agent debate prompt may be enough if your use cases are simple and you do not need structured verdicts.
• Workflow fit: Match the activation phrases to your own agent's vocabulary; if your team never says "battle this out", alias the triggers.

Why it is worth tracking

This skill addresses a real gap: agents are good at arguing but bad at extracting a decision. The forge-harness repository is early with 14 stars, so evaluate it before relying on it in production, but the structural pattern is worth tracking because it can be dropped into any agent that already supports session-inherited model context. You can inspect a live listing at https://www.openagentskill.com/skills/chrono-meta-deliberation to see how it behaves in practice. As the skill matures, watch for more persona jurors and stronger mediator heuristics — those are the signs it is becoming a durable workflow tool.