User reference / Run an experiment

3. Agents & rounds

Choose the participating models, refinement rounds and voters for this case.

Reviewed 2026-09-21Process executor 0.1.5

Define the participants for this case

llm_configs is an ordered list of model configurations. Agent slot 1 uses entry 1, slot 2 uses entry 2, and so on; when there are more agents than entries, assignment wraps around. The same slot-to-model mapping is reused for every refinement and for voting. The analyst has its own configuration.

EXAMPLE / REFERENCE

{
  "base": "slow-thinker.base.json",
  "label": "comparison 3x4",
  "num_proposing_agents": 3,
  "num_refinement_rounds": 4,
  "num_voting_agents": 3,
  "max_steps_in_summaries": 0
}

Save this example beside the base in config/. It reuses its models, keys and analyst, but requests five total proposal rounds: round 0 plus four refinements.

Council parameters

Documented in E 0.1.5

Parameter	Behavior
`num_proposing_agents`Reviewed E 0.1.5	Number of parallel proposal calls per round. Use a positive integer. It need not equal the number of model entries.
`num_refinement_rounds`Reviewed E 0.1.5	Number of additional rounds after the independent round 0. Zero means vote on the independent proposals. There is no early stop when titles stabilize.
`num_voting_agents`Reviewed E 0.1.5	Number of parallel voters, assigned from the same model list. Each valid vote chooses one final proposal.
`max_steps_in_summaries`Reviewed E 0.1.5	Maximum steps included when rendering proposals for refiners and voters. `0` includes all steps. A positive value truncates the step listing; it does not produce an intelligent summary.
`label`Reviewed E 0.1.5	Human-readable name shown in reports and histories. Labels such as “5x3” are conventions, not parsed instructions. The label also participates in configuration identity.
`analysis_llm`Reviewed E 0.1.5	Optional separate model configuration for post-run evaluation. If empty, the report command falls back to the first `llm_configs` entry when analysis is requested.

For N proposers, R refinements and V voters, the expected number of successful collaboration calls is N × (R + 1) + V. Three agents and two refinements produce nine proposals plus three votes. Five agents and three refinements produce twenty proposals plus five votes. Retry attempts and failed calls complicate actual billing; see cost accounting.

Model parameters and identity

Documented in E 0.1.5

Parameter	Meaning
`provider` / `model`Reviewed E 0.1.5	Select a registered adapter and the model identifier sent to that provider.
`name`Reviewed E 0.1.5	Optional stable human label. It must start with a letter and contain only letters, digits, `.`, `_` and `-`; council names must be unique.
`reasoning_effort`Reviewed E 0.1.5	Provider-specific reasoning control. It is not a comparable compute budget across providers.
`thinking_budget`Reviewed E 0.1.5	Provider-specific thinking token control. Its meaning, including zero, depends on the adapter.
`temperature`Reviewed E 0.1.5	Sampling setting only where the adapter sends it. Some reasoning adapters omit it.
`max_tokens`Reviewed E 0.1.5	Output token limit passed through the adapter; Anthropic may raise it to leave room above an extended-thinking budget.

The provider reference explains the effective settings, defaults and exceptions. Reports identify participants by the numeric suffix of IDs such as opus5_initial_1 and opus5_refine_1. “Own previous version” means the same slot in the preceding round. It does not mean the model retained conversational memory.

What the current architecture shares

Documented in E 0.1.5

In round 0, each agent receives the task independently. Each refinement receives all valid proposals from the preceding round, including author and model labels, but no special “your own proposal” marker. Agents use fresh calls. They may improve a proposal or adopt a different approach. In the final vote, author and model identities are removed from the proposal descriptions, although the text itself can still reveal stylistic clues.

The JSON parameters customize this architecture. They do not yet define arbitrary message routes, supervisors, critic roles or conditional loops. The post-run analyst is not an in-loop critic.

Continue with execution.

Source files used for this reference

Reviewed at 7361ea0b71a5.

← Documentation home Find a report element →