Skip to content

Role Orchestration on a Single Model

Invoke the same frozen small model in three distinct roles — summariser, agent, corrector — to roughly double task-goal completion on constrained hardware without additional training.

When This Applies

The technique targets a specific deployment constraint: an 8B-class model running on a single 24 GB GPU, where a larger model does not fit and additional training is not an option (McClendon et al., 2026). On unconstrained hardware, running a larger model once — or routing to heterogeneous models per role — is typically a stronger baseline. The result is a competitiveness claim for constrained deployments, not a general recommendation to stack roles on one model.

The Three Roles

The same frozen weights are invoked three times with different conditioning, each role presenting a different action space (McClendon et al., 2026):

Role Input Output Purpose
Summariser Full dialogue history Compressed summary preserving tokens, credentials, API responses Keep critical artifacts across long trajectories
Main agent Compressed context + task Tool calls and reasoning Drive the primary task loop
Corrector Agent's code output only (no conversation) Revised code Break repetitive failure loops 
graph LR
    H[Dialogue history] --> S[Summariser role<br>same weights]
    S -->|compressed context| A[Main agent role<br>same weights]
    A -->|code output| C[Corrector role<br>same weights]
    C -->|revised code| E[Environment]
    E --> H

The corrector receives code without conversation history. This isolation is the mechanism that breaks loops: without seeing the main agent's prior failed attempts, the corrector is not primed to repeat them.

Reported Results

On AppWorld (Trivedi et al., 2024) with Qwen3-8B on a 24 GB GPU (McClendon et al., 2026):

Configuration Baseline Scaffolded
FP16 (12K context) 5.4% 8.9%
AWQ 4-bit (32K context) 3.0% 5.9%
Difficulty-1 tasks (FP16) 15.8% 26.3%

The scaffolded 8B surpassed DeepSeek-Coder 33B Instruct (7.1% on FP16) from the original AppWorld evaluation (McClendon et al., 2026). Absolute completion remains far below GPT-4o's ~49% on AppWorld normal tasks (Trivedi et al., 2024) — the result is a small-model competitiveness gain, not a production threshold.

Why It Works

Each role presents a different action space to the same weights — different instructions, inputs, and output formats. The paper frames this as "a scaffolded policy over a frozen base model, three invocations of the same weights with different conditioning," drawing explicit connections to test-time compute scaling and action-space shaping in reinforcement learning (McClendon et al., 2026).

Diversity here comes from conditioning, not from heterogeneous weights. That distinguishes the technique from ensemble approaches and from dual-model critic agents that use a different model to avoid the self-correction blind spot measured at 64.5% across 14 LLMs.

Limits and Counter-Evidence

Same-weight review inherits shared failure modes. The corrector's isolation from conversation history mitigates loop repetition but does not eliminate the self-correction blind spot — the same weights still share representational biases (arXiv:2507.02778).

Role boundaries are prompt-level, not architectural. A frozen-weights scaffold cannot enforce role scope via permissions; agents frequently disobey role specifications when under-conditioned (Cemri et al., 2025).

The gain is task-dependent. The corrector role specifically addresses repetitive failure loops. Tasks that fail in novel ways each time show less benefit. Short-context tasks gain less from summarisation than long-trajectory tasks.

When This Backfires

  • Unconstrained hardware — running a larger model once, or routing different models per role as in cost-aware agent design, usually outperforms three invocations of one 8B model.
  • Weak base models — if the model cannot follow role-conditioned prompts consistently, the three invocations collapse to "same weights doing the same thing" with triple the latency.
  • No long dialogue history — the summariser role adds overhead when naive truncation or prompt caching would suffice.

Key Takeaways

  • Same frozen weights, three conditioning regimes (summarise, reason, correct) — a scaffolded policy, not an ensemble.
  • The corrector's isolation from dialogue history is the specific mechanism that breaks repetitive failure loops.
  • Scoped to constrained-hardware small-model deployments; larger models or heterogeneous role routing remain stronger on unconstrained hardware.
Last reviewed: 2026-05-27
Feedback