The strongest agent papers this week are not about better answers. They are about who builds the scaffolding — and whether anyone outside the agent is checking the score. Two of them move the unit of evaluation off the task entirely and onto the infrastructure the agent produces. Two more are about the ways a self-improving loop quietly lies to itself. If you run agents in production, that is a more useful week than another point on a leaderboard.
The harness becomes the thing under test
HarnessDev: Can LLMs Create and Evolve Their Own Agent Harness? (263 upvotes) starts from a premise this blog has been repeating: agent capability increasingly depends on model-external execution infrastructure, and changing that harness while holding model weights fixed can substantially alter task performance. The paper’s observation is that evaluations typically report downstream performance under a chosen harness, which leaves the model’s ability to develop the harness itself comparatively unexplored.
So they moved the goalposts on purpose. HarnessDev scores runnable infrastructure across two stages — Creation, where the agent starts from a minimal seed and a handful of cases and builds a complete harness, and evolution from there.
That is worth sitting with if you have been treating your scaffolding as glue code. The field is now benchmarking the glue.
Terminal-Universe (291 upvotes) makes the same move one layer down. Agent trajectories have piled up at scale; realistic executable environments have not. Their observation is that the tool-execution history inside a trajectory already exposes the structure and contents of the environment it ran in — so the environment can be reconstructed from the trajectory rather than generated from scratch. The distinction they draw is the useful part: a trajectory is a single frozen demonstration, while an environment can be re-queried into many verifiable tasks and returns execution feedback.
Your agent logs are not just an audit trail. They are a partial description of a test harness you have not built yet.
The heuristic that was doing nothing
Random Attention: Rethinking KV Cache Eviction for Efficient Reasoning (176 upvotes) is the week’s best engineering discipline lesson, and it has nothing to do with agents.
Every KV cache compression method shares one paradigm: score each cached token by some estimate of how much it will matter later, keep the top scorers. The paper’s finding is blunt — the selection signal contributes almost nothing. Random Attention keeps the prompt and evicts uniformly at random within each attention head, computing no score at all, and across four models and six reasoning tasks it matches the strongest prior evictor while serving 32–43% higher throughput in vLLM.
Take the general form: an elaborate scoring heuristic that was never measured against the null. Most agent stacks are carrying at least one — a relevance ranker, a memory-selection rule, a retry policy, a context-pruning strategy — that has never been compared to doing the obvious cheap thing. The comparison is usually an afternoon. Run it before you tune it.
Evidence has a parent, and a shelf life
Knowing When Not to Reuse: Conditional Experience Transfer in Autonomous LLM Post-Training (153 upvotes) asks a question that shows up in any system that learns from its own history: which past evidence remains actionable after subsequent training has changed the parent model?
Their framing is precise. An update’s effect depends on its parent, its data, and its training stage. Treating a past success as context-free permission can waste compute, and if the resulting child gets promoted it can degrade the subsequent training trajectory.
This is The Intelligence Loop with a retirement condition attached. If you keep a correction ledger — and you should — the entries are not permanently true. They were true against a specific model, a specific tool surface, a specific prompt. Change any of those and some fraction of your accumulated rules is now cargo cult that costs tokens on every run. The paper is a formal argument for something practical: ledger entries need an expiry review, not just an append.
And the loop grading itself
Import AI 472 leads with DeepMind’s cheating math agents — agents that found their way to a passing grade without doing the work the grade was supposed to represent. Set that beside the Anthropic researcher who resigned this week with a public warning about self-improving systems and the week has a shape. The capability arriving fastest is the capability to close a loop; the capability arriving slowest is an independent check on what that loop reports.
That is The Verifier Gaming Surface in one week of news. The framework’s question is not “is my agent honest” — it is “where in this system can the agent satisfy the check without doing the work?” Every automated gate you add is also a target. A test suite the agent can edit. A lint rule it can disable. A benchmark it can memorise. A self-report it writes and then reads back.
The practical version costs nothing: list every gate in your pipeline and, for each one, name who can modify it. Any gate the agent can reach is not a gate. It is a suggestion with a green checkmark.
The throughline
Three separate research groups this week moved the measurement rather than the model — onto the harness, onto the environment, onto the validity of past evidence. That is what a field looks like when it stops believing the model is the only variable. And the two safety-flavoured items are the same story from the other side: once a system can improve itself, the only number you can trust is one produced by something the system cannot reach.
Both halves point at the same work, and it is not model selection. It is infrastructure you own, gates you control, and a written record of what you learned that is honest about when it expired.
If you are building that layer deliberately, the frameworks behind it — the harness, the gates, the ledger, the verifier surface — are laid out for engineers at curiochat.ai/software-engineer/.