Most agent research focuses on fixed harnesses or one-off improvements; this paper makes the opposite choice explicit: let the harness itself be a first-class, self-developing artifact. The core insight is that an agent can schedule and review its own structural changes (code, prompts, tools, workflows) so those changes become the runtime for future work, producing long-lived evolutionary lineages rather than disposable experiments. That shift changes evaluation, safety, and what it means to iterate on agent designs.
Key Findings
- Two complementary evolution modes: recursive free evolution (where improvement is treated as a task and can chain into further evolutions) and experience-driven core evolution (where real work and social interaction surface bugs and structural fixes). This means the agent can both proactively search for improvements and react to operational failures.
- Empirical results show competitiveness on multiple coding benchmarks: an Opus 5 run achieved 86.74% on Terminal-Bench 2.1 and 90.69% on OSWorld-Verified; a five-rollout CL-Bench campaign reached a normalized reward of 0.2301. These numbers position the system at or above prior state of the art for the reported setups.
- Long-lived live deployment: "Hope" is a 161-day documented free-evolution deployment spanning seven interaction surfaces; humans propose changes but the agent chooses which to apply, illustrating sustained autonomous development under social input.
- Safety and evaluation practices: because the agent can rewrite its own code and swap model APIs, the paper separates live evolutionary lineages from frozen-snapshot benchmarks. Operational guardrails and governance are treated as primary design problems rather than afterthoughts.
Who It's For and Tradeoffs
Great fit if you research autonomous coding agents, long-horizon agent design, or agent governance: the paper supplies both a conceptual framework (how to let a harness evolve) and empirical evidence that such evolution can produce measurable gains. It is valuable for teams wanting reproducible benchmark comparisons while running parallel live evolution experiments. Look elsewhere if you need a ready-to-deploy production agent today: the approach raises distinct safety, reproducibility, and maintenance challenges (agents rewriting core code, changing dependencies, or switching model providers). Benchmarks require frozen snapshots to remain comparable, and live lineages demand human-in-the-loop governance and monitoring to manage drift and emergent behaviors.
Where It Fits
This work sits between bench-focused agent evaluation (short-run, frozen experiments) and speculative autonomous-development research. It provides a pattern for projects that want to combine rigorous benchmarking with continuous live improvement, and it foregrounds governance practices that many agent architectures currently ignore.
How It Works (brief)
The system treats evolution cycles as formal tasks: propose structural changes, submit for (human or agent) review, merge reviewed commits into the harness, and run new workloads on the updated runtime. The harness records durable memory and history across restarts, delegates to specialist agents when appropriate, and uses frozen snapshots for benchmark campaigns while allowing a separate live lineage to continue evolving.