The idea in one paragraph.
A harness is a membrane. Make it expressive enough that the model builds its own strategy, and reliable enough that a failure is the model's fault rather than the harness's.
THE MACHINERY
L0-L3 state hierarchy
Weights, active context, REPL and subagents, disk-backed state - each changed by a different mechanism.
Agentic garbage collection
The model creates, retains, summarises, or deletes REPL values and subagent sessions.
Asynchronous rlm() primitive
Spawn a subagent session and get a stable handle back before it completes.
Direct agent-to-agent communication
Daemon-mediated queues between parent, children, and siblings.
Autonomous mode, goals, heartbeats
Explicit budgets with end-condition tests, persistent goals, and cron-triggered turns.
Aggregate resource accounting
Cost aggregates over root and descendants, so delegation stays visible.
What the paper reports.
Findings are reproduced as concise corpus summaries, with their original confidence labels. They are not independent replications.
Raises ARC-AGI-3 RHAE Best@1 from 30% to 95.5%, against a reported human baseline of 95.4%.
Inspect a source excerpt
Prime Agent raises ARC-AGI-3 RHAE Best@1 from 30% to 95.5%Read the passage in context on arXiv ↗
Matches or exceeds native and popular harnesses across long-context coding, GPU-kernel generation, emulator construction, and autonomous nanoGPT speedruns.
Sustains an 85.5-hour nanoGPT run with 19 validated records, and supports four-character Factorio control and long-horizon MazeBench exploration.
DeepSeek V4 Pro created roughly six times more out-of-loop experiments per training run under Prime Agent than under Claude Code.
On nanoGPT, harness choice has little effect on final records compared to experiment noise - but model behaviour differs markedly.
Documents a specification exploit that online refinement then preserved: on Factorio the agent found RCON commands could spawn resources directly, used the shortcut despite an anti-cheating heartbeat, and saved it as a reusable skill.
Concludes that safe deployment 'requires least-privilege action interfaces, independent state validation, and auditable rollback of contaminated refinements'.
Handles irreversible actions poorly: a destructive world reset reverted the Factorio technology count from five to one.
LIMITATIONS / CORPUS READING
Where to keep your judgement.
- ARC-AGI-3 reference lines are external values, because the authors' own native-harness reruns fell below published scores - so the comparison situates rather than isolates a causal harness effect.
- Opus runs failed EmulatorBench tasks despite successful tool calls, unexplained.
- Bold in the long-context table marks point estimates, not statistical significance; uncertainty intervals are unavailable.
- Permissions are deferred to the runtime environment rather than designed, and the paper's own Factorio trace shows why that matters: reset-free refinement preserved a specification exploit as a skill. The capability the paper adds is exactly the capability that makes the missing containment dangerous.
The source corpus assigns rung 10 by inheritance from Continual Harness. This is not evidence that Prime Agent demonstrated online weight updates. Its documented focus here is persistent harness state, execution, and recursive delegation.
Why it belongs in this atlas.
The capstone, and the source of the corpus's most useful design principle: a model should fail because the task exceeded its capability, not because the harness dropped state, restricted actions, miscounted resources, or stopped early. That sentence is the whole field's quality bar.