Technical Report · 2026

HarnessX

A Composable, Adaptive, and Evolvable Agent Harness Foundry

Darwin Agent Team 5 benchmarks 15 model-benchmark configurations

Read paper GitHub
HarnessX composes agent behavior, adapts the harness from agent runs, and evolves the model from improved trajectories
HarnessX unifies composition, adaptation, and evolution in one runtime loop.

Overview

Treat the harness as a first-class interface.

An agent is more than a model call. Its harness decides how a task is represented, what context and memory persist between steps, which tools can be invoked, where side effects occur, how outcomes are scored, and which traces are retained. HarnessX makes this runtime interface a typed object that can be composed, audited, and improved from execution feedback.

01

Compose

Swap behavior across nine orthogonal dimensions without rewriting the agent.

02

Adapt

Turn traces into bounded harness edits, then test and gate every candidate.

03

Evolve

Convert reward-annotated trajectories into training signal for a better model.

Representative edit surfaces

One harness, nine edit surfaces.

An agent is represented as H = (M, C): model configuration M chooses role and fallback models, while C owns the behavior pipeline and shared runtime slots. Processors attach to lifecycle hooks, so a single behavior can be inserted, replaced, or removed without rebuilding the agent.

Typed interface: eight lifecycle hooks, same event type in and out, deterministic ordering, singleton groups, and serializable configurations.

Compose

  • D1 Model selection
  • D2 Context assembly
  • D3 Memory management
  • D4 Tool ecosystem
  • D5 Execution environment

Adapt

  • D6 Evaluation and reward
  • D7 Control and safety

Evolve

  • D8 Observability
  • D9 Training bridge

AEGIS

The harness learns from its own traces.

AEGIS treats harness adaptation as reinforcement learning in symbolic space: the current harness and trace store form the state, a typed code-level edit is the action, and execution traces plus verifier scores provide feedback. One meta-agent selectively invokes four stages; only verified improvements become the next harness.

AEGIS workflow from HarnessX composition and runtime traces through Digester, Planner, Evolver, and Critic
AEGIS evolves typed harness configurations with an explicit ship-or-reject gate.

Digester

Compresses roughly 10M raw trace tokens from one GAIA pass@2 iteration → ~10K structured summaries with cross-round history.

Planner

Builds an adaptation landscape covering failed tasks, implicated components, and untried prompt, tool, processor, or config edits.

Evolver

Emits typed builder candidates with a change manifest; new processor code also carries a smoke test.

Critic + gate

Checks evidence and non-local risks, then verifies normalization, build checks, and regressions before shipping.

Every round: run the current harness, retain complete traces, test candidates on the same adaptation batch, and archive rejected edits with a reason.

Seesaw constraint: a candidate cannot regress a previously solved task. When tasks conflict, variant isolation can route them to separate harness versions instead of discarding a useful local fix.

Results

Harness evolution improves 14 of 15 configurations.

The evaluation spans GAIA, ALFWorld, WebShop, tau3-Bench, and SWE-bench Verified across three model families. The largest gains appear where the initial model has the most behavioral gaps; reported peak scores use the same task sets used for evolution.

5 benchmarks GAIA 103 (exact match), ALFWorld 134 (goal completion), WebShop 100 (attribute match), tau3-Bench 3 domains (rule compliance), SWE-bench Verified 55 (patch resolution)
3 task agents Claude Sonnet 4.6, GPT-5.4, and Qwen3.5-9B; Opus 4.6 runs the meta-agent
Up to 15 rounds Full task set evaluated each round, with early stop after three idle rounds
Pass@2 Two independent attempts; either success counts. No held-out generalization is reported.
+14.5% average absolute gain across all 15 configurations
+44.0% largest gain, on ALFWorld with Qwen3.5-9B
14 / 15 model-benchmark configurations improved

Representative runs

Peak pass@2, percentage points

BenchmarkTask agentInitialPeakGain
ALFWorldQwen3.5-9B53.097.0+44.0
GAIAQwen3.5-9B20.337.4+17.1
WebShopGPT-5.455.073.0+18.0
SWE-bench VerifiedGPT-5.445.563.6+18.2
tau3-BenchGPT-5.476.290.7+14.5
ALFWorldQwen3.5-9B 53.0 → 97.0 +44.0
GAIAQwen3.5-9B 20.3 → 37.4 +17.1
WebShopGPT-5.4 55.0 → 73.0 +18.0
SWE-bench VerifiedGPT-5.4 45.5 → 63.6 +18.2
tau3-BenchGPT-5.4 76.2 → 90.7 +14.5

Heterogeneous tasks: GAIA with GPT-5.4 stagnates under one global harness (73.8% peak, 49.5% final). Variant isolation reaches 87.4% with peak = final, +13.6 points over the static baseline, while using 107.8M vs 143.7M tokens.

GAIA-103 harness evolution trajectory with per-round pass@2, best-so-far, and annotated failure analysis
GAIA-103 evolution trajectory with annotated failure patterns from the technical report.

Harness-model co-evolution

The same trajectories update both sides.

Each iteration rolls out the pair (Mt, Ht), scores traces with a fixed verifier, and appends them with their harness version to a FIFO replay buffer. AEGIS makes the discrete harness edit; cross-harness GRPO groups trajectories by task across harness versions and learns from their relative rewards. The next round starts with (Mt+1, Ht+1) without extra environment rollouts for the model update.

HarnessX co-evolution workflow: shared task traces feed AEGIS harness evolution and cross-harness GRPO model updates
One replay buffer supplies evidence for harness search and training signal for the model.

Measured gain

Search and learning reinforce each other.

On GAIA (103 tasks) and WebShop (100 tasks) with Qwen3.5-9B, a four-round FIFO window lets both updates learn from the same recent trajectories. Interleaving cross-harness GRPO with AEGIS raises peak success by +4.7% on average over harness-only evolution.

GAIA: 37.4% to 41.7% (+4.3%). WebShop: 49.0% to 54.0% (+5.0%). The gap remains at the final round.

AEGIS and model-harness co-evolution success rates on GAIA and WebShop
Co-evolution raises peak success on both evaluated benchmarks.
All 15 model-benchmark configurations (peak pass@2)
BenchmarkTask agentInitialPeakGain
ALFWorldClaude Sonnet 4.683.694.8+11.2
ALFWorldGPT-5.476.997.8+20.9
ALFWorldQwen3.5-9B53.097.0+44.0
WebShopClaude Sonnet 4.660.076.0+16.0
WebShopGPT-5.455.073.0+18.0
WebShopQwen3.5-9B36.049.0+13.0
GAIAClaude Sonnet 4.673.883.5+9.7
GAIAGPT-5.473.873.80.0
GAIAQwen3.5-9B20.337.4+17.1
SWE-bench VerifiedClaude Sonnet 4.676.487.3+10.9
SWE-bench VerifiedGPT-5.445.563.6+18.2
SWE-bench VerifiedQwen3.5-9B23.641.8+18.2
tau3-BenchClaude Sonnet 4.689.695.0+5.4
tau3-BenchGPT-5.476.290.7+14.5
tau3-BenchQwen3.5-9B93.594.6+1.1
What the trace analysis revealed

Reward hacking

GAIA/Sonnet rose 74.8% to 79.6% through verifier-format regularities; trace analysis led to a cross-check guard two rounds later.

Catastrophic forgetting

A sixth same-type reminder on tau3-Bench Telecom drove compliance from 94.7% to 80.7% at R7, then recovered by R9.

Under-exploration

On ALFWorld/Sonnet, prompt edits yielded less than 1% per round from R4-R7 while ship-prediction accuracy fell from 80% to 0%.

Citation

Read the full report.

The report contains the complete method, evaluation protocol, ablations, and failure analysis.

Open report Repository 
@techreport{darwinagent2026harnessx,
  title  = {HarnessX: A Composable, Adaptive, and
            Evolvable Agent Harness Foundry},
  author = {Darwin Agent Team},
  year   = {2026},
  url    = {https://github.com/Darwin-Agent/HarnessX}
}