Subagent Assignment Matrix (Empirical Local Bench)

Updated: 2026-02-23
Source benchmark: generated/model_profiles_2026-02-23.json

Speed Summary (lower is better)

  • qwen2.5:14b avg: 13.32s (fastest)
  • mistral-small3.2:24b-instruct-2506-q4_K_M avg: 24.78s
  • qwen2.5-coder:32b-instruct-q3_K_L avg: 36.36s

Routing Policy (Local-first + Codex 5.3 for critical tasks)

Hard rule:

  • Any local-model task must route through skills/ollama-agent-queue (never direct ollama run in task scripts).
  • If local queue fails/timeouts, escalate to API fallback.

Default model for most subagents

  • qwen2.5:14b
  • Use for: operations, summaries, drafting, monitoring, report generation, first-pass analysis

Coding-heavy / high-complexity code

  • qwen2.5-coder:32b-instruct-q3_K_L
  • Use for: root-cause debugging, non-trivial refactors, architecture/code-level reasoning
  • Guardrail: only invoke when task is explicitly code-heavy or qwen14b fails first pass

Quality second-opinion / rewrite pass

  • mistral-small3.2:24b-instruct-2506-q4_K_M
  • Use for: alternative framing, concise rewrite, consistency check

Practical Assignment Matrix

Task Type Primary Fallback 1 Fallback 2 Notes
Daily brief / ops summary qwen2.5:14b (via queue) mistral-small3.2:24b (via queue) Codex 5.3 (Plus/API) Optimize for speed
Research scan + synthesis qwen2.5:14b (via queue) mistral-small3.2:24b (via queue) Codex 5.3 (Plus/API) Escalate for strategic depth
Social copy draft batch qwen2.5:14b (via queue) mistral-small3.2:24b (via queue) gpt-4o (API) Use API for final polish
Browser automation planning qwen2.5:14b (via queue) qwen-coder:32b (via queue) Codex 5.3 (Plus/API) Coding model when selectors/scripts complex
Script writing / bugfix qwen-coder:32b (via queue) qwen2.5:14b (via queue) Codex 5.3 (Plus/API) Prefer coder model for non-trivial logic
Fast triage / inbox classification qwen2.5:14b (via queue) mistral-small3.2:24b (via queue) Throughput-first
Financial/compliance messaging qwen2.5:14b (via queue) Codex 5.3 (Plus/API) gpt-4o (API) Human confirmation still required
High-stakes strategy / external commitments Codex 5.3 (Plus/API) GPT-4o (API) Gemini (API) Director review required before send

Subagent Spawn Defaults

  • research-*, brief-*, monitor-*qwen2.5:14b
  • code-*, debug-*, automation-*qwen2.5-coder:32b-instruct-q3_K_L
  • rewrite-*, second-pass-*mistral-small3.2:24b-instruct-2506-q4_K_M

Escalation Triggers

Escalate from local to API model only if one or more are true:

  1. Local output fails quality after one revision loop.
  2. Time-critical blocker where local latency compounds delay.
  3. High-stakes external-facing deliverable requiring top polish.

Immediate Operational Change

For spawned subagents, set model explicitly per task type instead of defaulting every task to one model.