Ensemble

Ensemble consensus is a multi-model iterative refinement pattern. Three specialized models -- a Drafter, a Critic, and a Synthesizer -- collaborate in a loop to produce high-quality output that meets a configurable quality threshold.

---

Endpoint

POST /internal/agent/ensemble

This is an internal OIDC-authenticated endpoint on the agent executor service. It is not directly accessible to end users.

---

Concept

The ensemble pattern assigns three distinct roles to three different models, each operating at a different temperature to balance creativity and rigor:

                     +---------------------------+
                     |    Round N (N = 1..max)    |
                     +---------------------------+
                                |
                                v
                     +---------------------+
                     |      DRAFTER        |
                     |  (creative, high T) |
                     |  Generates output   |
                     |  + self-assessment   |
                     +---------------------+
                                |
                                v
                     +---------------------+
                     |       CRITIC        |
                     |  (rigorous, low T)  |
                     |  Scores dimensions  |
                     |  + verdict          |
                     +---------------------+
                                |
                                v
                     +---------------------+
                     |    SYNTHESIZER      |
                     |  (balanced, mid T)  |
                     |  Integrates feedback|
                     |  + decision         |
                     +---------------------+
                                |
                        +-------+-------+
                        |               |
                   APPROVED       ANOTHER_ROUND
                        |               |
                        v               v
                   Return output   Loop back to
                                   Drafter (round N+1)

Role Definitions

Role	Purpose	Temperature	Behavior
Drafter	Generate creative, comprehensive output	High (0.4--0.8)	Produces initial content. In rounds > 1, explicitly addresses previous feedback and critical issues. Includes a self-assessment section.
Critic	Evaluate output rigorously against quality dimensions	Low (0.2--0.3)	Scores each dimension 1--10. Provides composite score, verdict, critical issues, suggestions, and specific edits. Never generates content.
Synthesizer	Integrate feedback and decide whether to approve or loop	Medium (0.3--0.5)	Reviews drafter output and critic feedback. Applies non-destructive edits. Issues final APPROVED or ANOTHER_ROUND decision.

---

Request Body

{
  "user_id": "string",
  "task": "string",
  "drafter_model": "string",
  "critic_model": "string",
  "synthesizer_model": "string",
  "drafter_temperature": 0.8,
  "critic_temperature": 0.2,
  "synthesizer_temperature": 0.5,
  "max_rounds": 3,
  "approval_threshold": 8.0,
  "scoring_dimensions": [
    "accuracy",
    "completeness",
    "clarity",
    "actionability"
  ],
  "context": "Optional additional context injected into all prompts."
}

Parameter Reference

Parameter	Type	Required	Constraints	Description
user_id	string	Yes	--	Authenticated user ID for credit billing.
task	string	Yes	--	The task description provided to the Drafter.
drafter_model	string	Yes	Must be a valid model ID	Model used for the Drafter role.
critic_model	string	Yes	Must be a valid model ID	Model used for the Critic role.
synthesizer_model	string	Yes	Must be a valid model ID	Model used for the Synthesizer role.
drafter_temperature	float	No	0.0--1.0	Temperature for the Drafter. Default: 0.7.
critic_temperature	float	No	0.0--1.0	Temperature for the Critic. Default: 0.2.
synthesizer_temperature	float	No	0.0--1.0	Temperature for the Synthesizer. Default: 0.4.
max_rounds	integer	Yes	1--5	Maximum number of Drafter-Critic-Synthesizer loops.
approval_threshold	float	Yes	1.0--10.0	Minimum composite score required for approval.
scoring_dimensions	array	No	1--10 items	Dimensions the Critic evaluates. Default: ["accuracy", "completeness", "clarity", "actionability"].
context	string	No	--	Additional context injected into all model prompts.

---

Loop Mechanics

Round Execution

Each round proceeds through three phases:

Phase 1 -- Drafter generates output:

• Receives the original task and any context.

• If round > 1, also receives the Critic's feedback from the previous round, including critical issues and specific edit suggestions.

• Produces the output content plus a self-assessment section.

Phase 2 -- Critic evaluates output:

• Receives the Drafter's output and the original task.

• Scores each dimension in scoring_dimensions on a 1--10 scale.

• Produces a structured evaluation:

COMPOSITE SCORE: 7.2 / 10.0
VERDICT: NEEDS_REVISION

DIMENSION SCORES:
- accuracy: 8
- completeness: 6
- clarity: 8
- actionability: 6

CRITICAL ISSUES:
1. Missing competitive analysis section
2. Revenue projections lack supporting data

SUGGESTIONS:
- Add market size estimates with cited sources
- Include risk mitigation strategies

SPECIFIC EDITS:
- Section 3, paragraph 2: Replace "significant growth" with quantified projections
- Add new subsection: "Competitive Landscape" between sections 2 and 3

Phase 3 -- Synthesizer decides:

• Receives the Drafter's output and the Critic's evaluation.

• Integrates non-destructive edits where possible.

• Issues a decision:

  • DECISION: APPROVED -- output meets quality threshold, return to caller.

  • DECISION: ANOTHER_ROUND -- output needs revision, loop back to Drafter.

Approval Logic

The Synthesizer approves the output when both conditions are met:

1. The Critic's composite score is >= approval_threshold.

2. The Critic's verdict contains zero critical issues.

If either condition fails and rounds remain, the Synthesizer issues ANOTHER_ROUND.

Final Round Behavior

On the final round (round == max_rounds), the Synthesizer always auto-approves regardless of score. If unresolved critical issues remain, the Synthesizer appends [ESCALATE] notes to the output, identifying issues that could not be fully resolved within the allotted rounds.

---

Active Ensemble Configurations

The platform ships with five pre-configured ensemble profiles optimized for common use cases:

Ensemble	Drafter	Critic	Synthesizer	Max Rounds	Threshold
Content	claude-sonnet (T=0.8)	gpt-5.2 (T=0.2)	gemini-3.1-pro (T=0.5)	3	8.0
Research	gpt-5.2 (T=0.4)	claude-opus (T=0.2)	gemini-3.1-pro (T=0.3)	2	8.5
Strategy	claude-opus (T=0.6)	gpt-5.2 (T=0.2)	claude-sonnet (T=0.5)	2	8.0
Code Review	qwen3-coder (T=0.3)	claude-sonnet (T=0.2)	gpt-5.2 (T=0.2)	3	9.0
Comms/PR	claude-sonnet (T=0.6)	gemini-3.1-pro (T=0.2)	gpt-5.2 (T=0.4)	2	8.5

Configuration Notes

• Content: High-creativity drafter for marketing and editorial work. Rigorous fact-checking critic. Three rounds allow iterative refinement.

• Research: Moderate-temperature drafter for analytical tasks. Highest threshold (8.5) ensures factual rigor.

• Strategy: Claude Opus drafts strategic documents. Cross-vendor critic provides independent assessment.

• Code Review: Lowest temperatures across all roles. Highest threshold (9.0) enforces strict correctness. Three rounds for thorough iteration.

• Comms/PR: Optimized for professional communications. Two rounds balance quality with speed.

---

Plan Limits

Plan	Max Rounds Allowed	Notes
Free	Not available	Ensemble is not available on the Free plan.
Guru	2	max_rounds capped at 2 regardless of request value.
Pro	3	max_rounds capped at 3. Full access to all configurations.

Requests exceeding the plan's round limit will have max_rounds silently clamped to the plan maximum. The response includes the effective max_rounds used.

---

Example Request

POST /internal/agent/ensemble
Authorization: Bearer <OIDC token>

{
  "user_id": "uid_abc123",
  "task": "Write a 1500-word blog post on the impact of edge computing on IoT latency, targeting a technical audience familiar with cloud infrastructure.",
  "drafter_model": "claude-sonnet",
  "critic_model": "gpt-5.2",
  "synthesizer_model": "gemini-3.1-pro",
  "drafter_temperature": 0.8,
  "critic_temperature": 0.2,
  "synthesizer_temperature": 0.5,
  "max_rounds": 3,
  "approval_threshold": 8.0,
  "scoring_dimensions": [
    "accuracy",
    "completeness",
    "clarity",
    "technical_depth",
    "actionability"
  ],
  "context": "Target publication: company engineering blog. Audience: senior engineers and architects."
}

Example Response

{
  "status": "completed",
  "ensemble_id": "ens_7f3a9c2e",
  "rounds_used": 2,
  "max_rounds": 3,
  "final_score": 8.6,
  "approval_threshold": 8.0,
  "decision": "APPROVED",
  "scoring_dimensions": {
    "accuracy": 9,
    "completeness": 8,
    "clarity": 9,
    "technical_depth": 8,
    "actionability": 9
  },
  "output": "# Edge Computing and IoT Latency: A Technical Deep Dive\n\n...(full blog post content)...",
  "escalations": [],
  "round_history": [
    {
      "round": 1,
      "drafter_model": "claude-sonnet",
      "critic_score": 6.8,
      "critic_verdict": "NEEDS_REVISION",
      "critical_issues": [
        "Missing latency benchmarks comparing edge vs cloud",
        "No mention of 5G integration considerations"
      ],
      "synthesizer_decision": "ANOTHER_ROUND"
    },
    {
      "round": 2,
      "drafter_model": "claude-sonnet",
      "critic_score": 8.6,
      "critic_verdict": "EXCELLENT",
      "critical_issues": [],
      "synthesizer_decision": "APPROVED"
    }
  ],
  "credits_used": 342,
  "model_calls": [
    { "model": "claude-sonnet", "role": "drafter", "round": 1, "credits": 68 },
    { "model": "gpt-5.2", "role": "critic", "round": 1, "credits": 45 },
    { "model": "gemini-3.1-pro", "role": "synthesizer", "round": 1, "credits": 52 },
    { "model": "claude-sonnet", "role": "drafter", "round": 2, "credits": 74 },
    { "model": "gpt-5.2", "role": "critic", "round": 2, "credits": 48 },
    { "model": "gemini-3.1-pro", "role": "synthesizer", "round": 2, "credits": 55 }
  ]
}

---

Error Handling

Error	HTTP Status	Description
INSUFFICIENT_CREDITS	402	User does not have enough credits for the estimated ensemble cost.
PLAN_LIMIT_EXCEEDED	403	Requested max_rounds exceeds plan allowance (informational; rounds are clamped).
INVALID_MODEL	400	One or more specified models are not available in the model registry.
DRAFTER_FAILURE	502	The Drafter model returned an error during generation.
CRITIC_FAILURE	502	The Critic model returned an error during evaluation.
SYNTHESIZER_FAILURE	502	The Synthesizer model returned an error during integration.

On any model failure, the ensemble returns partial results including all completed rounds and the error detail.

---

Related

• Orchestration Overview

• Swarm Pipelines

• Credits System

• Models Reference