Learned Orchestration¶
Rather than coding the workflow, train the workflow.
What it is¶
Learned orchestration is a multi-agent design pattern where the orchestrator itself is a trained model rather than hand-written code. Instead of explicitly specifying which sub-agent runs in which order, a learned orchestrator discovers the optimal coordination strategy from data.
Sakana AI's Fugu (launched 22 June 2026) is the first production system to ship this pattern as an API product: a 7B-parameter conductor model that learns to route tasks to a pool of larger LLMs. Unlike its TRINITY predecessor, the shipped Fugu conductor does not assign roles — the selected model is always invoked as a plain worker, which narrows the coordination space and lowers orchestration latency. (Thinker / Worker / Verifier role assignment is TRINITY's design; see the TRINITY / Conductor architecture section below.)
How it differs from a hand-designed loop¶
| Hand-designed loop | Learned orchestrator |
|---|---|
You write the routing logic (if plan_step then call planner_agent) |
The conductor model learns routing from training data |
| Adding a new agent type requires a code change | The orchestrator adapts to new agents without code changes |
| Failure modes are the ones you anticipated | The system can discover non-obvious collaboration sequences |
| Version-controlled, auditable DAG | Behaviour emerges from weights — harder to audit |
The TRINITY / Conductor architecture (Sakana Fugu)¶
Two ICLR 2026 papers underpin Fugu's design:
TRINITY — an evolved coordinator assigns adaptive roles across a pool of models: - Thinker: deep reasoning about the problem - Worker: execution — code generation, API calls, file writes - Verifier: checks the worker's output before it returns to the caller
Conductor — uses reinforcement learning to discover "natural-language coordination strategies": the communication prompts between agents are themselves learned, not handcrafted.
Loop engineering implications¶
-
The harness as data, not code. If orchestration can be learned, the harness becomes a training target rather than a codebase. This is a fundamentally different philosophy from the ClaudeWarp / hand-designed approach (first-party: same maintainer as this KB).
-
Verification is still external. Even learned orchestrators need humans (or a separate verification loop) to validate final outputs — the system doesn't know what it doesn't know.
-
Provider-agnostic design matters more. A learned orchestrator dynamically selects the best model for each sub-task; coupling your loop to a single provider blocks this. See Failure Patterns → Provider lock-in.
A Second Target for Learning: the Harness Itself, Not Just the Orchestrator¶
Fugu learns which model does what (orchestrator-as-trained-model). A distinct but related idea trains the harness as a control layer while leaving the underlying LLM frozen: treat the scaffolding around the model — retry policy, tool sequencing, context management — as a learnable controller, trained via advantage-weighted regression on execution traces rather than hand-coded if/else logic. The paper introduces a Harness Maturity Score to make "how reliable is this harness's execution pattern" a measured quantity rather than a judgment call, distinguishing this from Harness Patterns' self-improving harnesses (which mutate the harness's code) by instead training a policy that sits on top of a fixed harness. (arXiv 2607.05458, Jul 2026.)
A periodization of the same trend, coined as three eras. A synthesis essay frames Bolt-On (ReAct-era prompted loops outrunning model capability) → Co-Training (RL trained inside the harness — e.g. codex-1's RL-on-real-tasks training — with models absorbing harness behavior like auto-compaction into their own weights) → a predicted Attention-Interface era, where every harness capability the model can absorb eventually gets deleted from the scaffold, leaving only human-facing controls (permissions, interruption policy) as the harness's remaining job. Testable and not yet observed: the essay predicts an "attention policy surface" ships alongside AGENTS.md within a year. (Dan McAteer / Latent Space, "The Evolution of the Agent Harness", Aug 2026.)
Training inside the harness you ship (Oct 2026). The Co-Training era in practice: Agent Lightning v1.0 (Microsoft Research; abstract read directly, an explainx.ai summary only pointed to it) names the paradigm harnessed agentic RL: the deploy-time harness, not the training engine, owns the environment loop, and the trainer sees only the LLM request-response pairs. The paper lists the resulting difficulties (retokenization, sample merging, advantage calculation, loss normalization, backend scheduling) and ships a roughly 3,500-line framework that supports arbitrary harnesses. Its self-reported coding result: Qwen3.5-9B on SWE-bench Verified from 41.8% to 56.4% with 6K training examples. At the other end of the spectrum, Harness² holds the weights fixed and optimises only the harness, while Evo² co-evolves weights and harness with one policy acting as both Proposer of harness edits and Reflector diagnosing their effect (abstract-level description). Together they bracket the design space: harness only, weights only (inside the real harness), or both.
A hand-designed analog of the same instinct — replan rather than restart on failure — appears in a large agent meta-harness's Replanning Loop: a GOAP-style A* planner decomposes a goal into an action sequence and replans specifically from the point of failure when state changes invalidate the current plan, instead of discarding progress and starting over. Paired with a Trust Loop that scores agent behavior (a weighted formula blending success rate, uptime, threat signals, and integrity) to upgrade or downgrade an agent's permission level over time. Treat the vendor's broader marketing claims (adoption numbers, "100+ agents") with skepticism — the concrete planning and trust-scoring mechanics are independently verifiable from the docs; the popularity claims are not. (ruvnet/ruflo, Jul 2026.)
Benchmarks (Fugu Ultra, June 2026)¶
| Benchmark | Fugu Ultra | Best frontier comparison |
|---|---|---|
| SWE Bench Pro | 73.7% | Claude Opus 4.8: 69.2% |
| LiveCodeBench | 93.2% | Gemini 3.1 Pro: 88.5% |
| GPQA-D | 95.5% | Gemini 3.1 Pro: 94.3% |
| Humanity's Last Exam | 50.0% | Claude Opus 4.8: 49.8% |
AutoResearch: a concrete loop engineering example¶
Sakana benchmarked Fugu-Ultra as the orchestrator on AutoResearch — Andrej Karpathy's autonomous ML-research benchmark, not a Sakana-built system — running 123 autonomous experiments against three anonymised frontier-model baselines — each iteration proposed a hypothesis, implemented a change, evaluated results, and fed findings into the next iteration. The loop ran unattended; humans reviewed summaries.
This is a direct example of the self-improvement loop pattern: a loop whose primary goal is to improve the loop itself, with a learned orchestrator deciding which experiments to run next. See The Loop Contract → self-discovery pattern.
Current state (June 2026)¶
- Sakana Fugu is available as an OpenAI-compatible API (swap endpoint, no SDK change) — two variants: Fugu (no fixed per-token price — billed at the standard rate of whichever underlying model is active, and never stacked when several agents run) and Fugu Ultra ($5 input / $30 output per 1M tokens, rising to $10/$45 above 272K context; also a $20–200/mo subscription covering both)
- Fugu Ultra is unavailable in EU/EEA pending GDPR compliance
- @steipete and others are skeptical about closed-source multi-model routing performance in practice
- No open-source learned orchestrator has matched Fugu Ultra's benchmark claims yet
- This pattern is early-stage — most practitioners still use hand-designed loops
Training Environments for Terminal Agents (Sept 2026)¶
Learned orchestration needs training environments that stay hard enough to be informative. Off-policy environment evolution keeps RL training environments near the model's learnable frontier, implemented via a loop-engineered multi-agent harness that evolves environments along three difficulty directions — a direct instance of "loop engineering" applied to the training pipeline rather than the deployed agent. Reports improving Qwen3.6-27B and Qwen3.6-35B-A3B by 14.4 and 18.0 points respectively on Terminal-Bench 2.1. (arXiv 2609.04128, "Environment Evolution for Terminal Agents", Sep 2026.)
Harness-RL — training the central controller, not the workers. A black-box RL framework with action-args decoupling trains a central controller that directs a multi-agent harness, rather than training the worker agents themselves — the RL counterpart to this doc's Sakana Fugu example above, but training the orchestrator role specifically. Reports 42.93/47.79 average F1 with Qwen models. (arXiv 2608.29641, "Harness-RL", Aug 2026.)
MemoryWalker — closing the train/inference mismatch from context compression. Harnesses compress context during a training rollout differently than they do at inference, producing a train-inference mismatch that quietly degrades a trained orchestrator's real-world performance. LogitTree and SDCC substantially close this gap across 7 benchmarks — relevant to any learned-orchestration setup that trains against a harness with its own compaction behavior (see Two Settings Tripled a Benchmark Score for why compaction settings alone can swing results this much). (arXiv 2609.00865, "MemoryWalker", Sep 2026.)
See also¶
- The Six Building Blocks — the hand-designed equivalent
- Fleet Engineering — managing many loops at enterprise scale
- Failure Patterns — what can go wrong in any orchestration approach