This week the harness stopped being scaffolding and became the thing under optimisation. Two of the most-upvoted agent papers automatically rewrite the prompts, control flow and context handling around a frozen model — one to cut cost roughly in half, one to show that the gains overfit unless you regularise them. Add a frontier price war and a new benchmark for agent judgment, and the message for anyone running agents is plain: the layer you own is now where most of the measurable movement is, and it needs the same held-out discipline as any model.
Self-improving harnesses overfit, and now there is a number for it
RRSI: Regularized Recursive Self-Improvement of Agent Harnesses (181 upvotes, code from Google Research) starts from an observation most teams running agent-tuning loops will recognise. Iteratively proposing and keeping harness edits is a form of recursive self-improvement at the system level, and it memorises: large in-distribution gains that “shrink or even vanish” out of distribution.
The fix borrows directly from ML regularisation. The proposer gets an annealed budget limiting how many edits one candidate can bundle; the selector adds a critic and a pruner that screen out benchmark-specific edits and remove changes that do nothing. Across eight benchmarks spanning coding, agentic workspace and engineering design tasks, RRSI reports up to 14.1 points in-distribution, up to 4.7 points on five out-of-distribution benchmarks, and 30% fewer policy tokens than unregularised evolution.
This is The Intelligence Loop with its failure mode named. Turning failures into permanent capability works — but a loop that only ever sees its own training tasks is fitting to them, and the gap between 14.1 and 4.7 is what that costs. If you run an improvement loop over your own agent’s rules or skills, hold out tasks it never learns from and score against those. The ledger that feeds the loop is covered in the self-improving workflow; the held-out set is what keeps it honest.
Half the tokens, same results, no new model
SoL-Pi (129 upvotes, NVIDIA) scales auto-research loops across many environments, keeping only harness mechanisms that transfer. Four survived selection: action execution, context compaction, observation handling, and delegated reading. On the 51-task EdgeBench evaluation it performs comparably to the Pi harness on both GPT-5.6 Sol and Opus 5 while cutting recorded token traffic by 44.7–49.0% and API cost by about a third — an estimated $8.75–$13.50 per hour saved against the native Codex and Claude Code harnesses.
Read it beside RRSI. SoL-Pi’s selection across diverse environments is doing the regularisation job RRSI does explicitly: a mechanism that only helps on one task never makes the cut. That is also The Token Budget Framework’s core claim, measured: cost is an engineering property of the harness, and compaction plus disciplined observation handling are where the spend goes.
The price war, and what it does not fix
Tuesday’s releases, summarised with prices by Simon Willison: Claude Opus 5.5 at $4 / $20 per million input/output tokens (down 20% on both, with cache reads down 60%), and an hour later GPT-6 Sol at $2 / $10 and GPT-6 Luna at $0.10 / $0.50, each half the price of its GPT-5.6 predecessor. Ars Technica’s framing is fair: a little more capability for a lot less money.
One detail from Willison belongs in every routing policy. At its “max” reasoning level, Opus 5.5 “over-thinks to breaking point,” exhausting its 128,000-token output limit on a single SVG and costing $2.56 per failed attempt. A price cut lowers the rate. It does nothing about a request that spends the whole budget and returns nothing. Effort level is a cost control, and the cheapest setting that clears your bar is still the default — route down, not up.
Agents are measurably weak at the forks
The Tasteful Agent (115 upvotes) measures something end-to-end benchmarks cannot see: whether an agent picks the right branch at the decision points that shape a long run. Taste-Bench mines those forks automatically from parallel attempts and from detours within single trajectories. The best frontier model answers 59.7% correctly. Accuracy drops further when the deciding evidence only appears later in the trajectory, and larger reasoning budgets barely help. The encouraging part: distilling an outcome-aware teacher’s judgments into a student improves both fork decisions and end-to-end success on held-out SWE-bench Pro tasks.
For now, the practical reading is The Human Gate Protocol. Gates belong where the run commits to a direction — which hypothesis, which implementation to build on — not only where it touches something destructive. A 59.7% decision is one you review before the agent builds a day’s work on top of it.
Benchmarks you can re-run
UK AISI is adopting EvalEval’s infrastructure to publish results as Evaluation Cards on a shared schema, starting with verified results and configuration details for five benchmarks — HealthBench, FrontierMath, Humanity’s Last Exam, SWE-Bench Pro and Terminal-Bench 2.0 — across six frontier models. The problem they name is the one every model-selection decision runs into: results reported in many formats, without enough information to reproduce them, from runs too expensive to repeat.
This is The Agent Observability Framework applied to evaluation itself. If a number is going to drive a routing decision, you want its configuration and transcripts, not a leaderboard cell.
The throughline
Every item this week moves the same line. The model is becoming a commodity input with a falling price; the harness, the gates and the evaluation are where reliability and cost are decided. RRSI adds the warning that comes with owning that layer: once your system improves itself, it can fit your tests instead of your work, and only a held-out set will tell you which one happened.
The frameworks for that layer — the improvement loop, the token budget, the human gate, observability — are laid out for engineers at curiochat.ai/software-engineer/.