Production agents can now learn from their own mistakes without retraining

What's new in AI research, what's gaining acceptance, and what newer evidence has overturned.

If you run a customer-facing agent, you have watched it repeat the same error pattern because your policy changed since the training data was collected. This week two production systems showed how to close that loop, and a training fix arrived for a reward method that quietly collapses diversity.

A team at Meituan's Keeta unit cut reply latency 28% and cost 71% by replacing a frontier proprietary model with a 31B in-house agent that improves itself nightly. Researchers at Princeton and Meta showed that executable strategy blueprints discovered on one model transfer to others with no code changes, cutting inference cost 39% while matching accuracy. And a group at Edinburgh proved that a common reinforcement-learning trick for reasoning models, GRPO, turns group rewards into a zero-sum tournament that punishes diversity, while their fix restores the true gradient with only two samples per prompt.

The production loop that stays honest to old policy versions

The hardest part of teaching agents from live traffic is that your business rules change while the failures you want to learn from happened under older rules. A team led by Yanbei Jiang at Keeta, Meituan, with colleagues at Melbourne and Fudan, built CARE, a recursive self-improvement system that trains on a mixture of failures from different policy versions without rewriting any of them to the current rulebook.

The key move is policy-conditioned learning. Each repair tuple is verified against the policy in force at incident time, not today's policy, and the training mix deliberately includes failures from multiple versions. A separate simulation test, frozen from the current policy, gates promotion. The result: a 31B model that started at 6.42% inspection bad-case rate reached 1.97% after five iterations, with median reply latency down 27.7% and cost per service order down 70.5% at full replacement. User satisfaction stayed parity with the incumbent GPT-5.4.

The training method matters. Standard supervised fine-tuning and direct preference optimization both collapsed on small repair batches. The team instead uses attribution-guided on-policy self-distillation, where a frozen copy of the same model acts as teacher with privileged hindsight appended to its system message, and the student learns from its own attempts truncated at the first error. Forward KL divergence, truncated to the teacher's top-32 tokens and floored at zero, kept training stable where alternatives failed.

The number that matters. Cost per service order fell 70.5% at full replacement, with model serving cost alone down 92.8%.

That cost reduction is what makes the loop sustainable. The queue empties faster now, and the agent pays for its own inspection infrastructure.

CARE: Grounded Recursive Self-Improvement for Production Customer-Service Agents - full text

Executable blueprints that travel between models

A separate team led by Taiming Lu at Princeton and Qi Xu at Meta took a different angle on the same problem: instead of improving one model's weights, improve the strategy it runs. Strategist evolves reasoning as executable Python blueprints with a single interface, execute(llm, task) -> str, stored in a shared library with measurement cards and experience notes.

The meta-agent searches over this library for ten iterations, retrieving relevant blueprints via hybrid search and proposing new ones that compose or reuse prior work. On 30 benchmarks, it matched the strongest hand-tuned baseline on 24 and tied on 2, with weighted inference cost down 39%. The ablation is stark: disabling the library drops accuracy 13.3 points and raises cost, while zero-shot selection without library or iteration crashes or hits 39.3 points below full.

More striking is cross-model transfer. Blueprints discovered on Qwen3 ran unchanged on GPT-5 and Gemini-2.5, winning on 7 of 10 and 8 of 10 benchmarks respectively, with aggregate cost down 35% and 40%. The tier-to-model mapping is fixed; the strategy code needs no modification.

The number that matters. One-shot retrieval from the final library, with no search iterations, stays within 2 accuracy points of the strongest baseline on all 10 held-out benchmarks and beats it on 8.

That means a builder can amortize search cost across deployments. The library pays for itself once, then every new task gets a near-optimal strategy at 0.77x baseline cost.

Strategist: Evolving Reasoning with Agents - full text

GRPO's hidden tournament structure

A training method widely used for reasoning models has a flaw that only shows up when rewards depend on the group. Junyi Gao at Edinburgh, with Yu Shi at Toronto and Pingzhao Hu at Western, proved that GRPO's group-relative baseline changes the optimization objective: it induces a zero-sum tournament where each response is trained to beat its groupmates, with externality weight α = −1/(K−1).

The effect is severe. On GlobalOpinionQA, GRPO collapses onto the majority answer and ends further from human opinion than the untrained model. On NoveltyBench, diversity-weighted rewards fail to prevent collapse: distinct answers fall from 4.33 to 2.65 while first-sample quality rises, a classic diversity-utility tradeoff hidden inside a method sold as improving both.

Their fix, Team-GRPO, sets α = +1 and restores the true gradient. It works with only two samples per prompt, credits each response with K times its change in group mean reward, and matches the target distribution where GRPO diverges. At matched KL divergence, Team-GRPO still leads by +0.84 in utility on NoveltyBench.

The number that matters. GRPO's error on distributional alignment reaches 0.491, more than double the base model's 0.197, while Team-GRPO reaches 0.081.

That gap is the difference between a model that reflects human disagreement and one that erases it. For any builder using group rewards, the sign of the externality term is not a detail; it is the whole objective.

When Rewards Depend on the Group, GRPO Plays a Tournament - full text

Verified guardrails that prove their own repairs

Sreehari Sankar and colleagues at Case Western Reserve built LeanGuard, a tool-use guardrail with formally verified admission logic. The evaluator is proved equivalent to independent LeanLTL temporal-logic semantics, and verification rejects any theorem depending on nonstandard axioms.

On the τ²-bench, it prevented all 33 policy-forbidden tool calls that unguarded agents executed, with task success preserved at 70.9% versus 72.7% unguarded when feedback is provided. Direct prompt-injection attack success on DTap fell from 57.9% to 5.9%. A proof-guided repair tactic, using Lean metaprogramming, proved admissible edits for 34.9% of argument-value denials in a median of 2 ms without involving the model at all.

The feedback mechanism matters. Bare denial drops task success to 68.9%. Providing a formal decision record plus natural-language hint, and letting the model fix and reissue, recovers most of the gap. Repeated blocked calls fall from 983 to 719, and recoveries rise from 34 to 48.

LEANGUARD: Lean-Verified Guardrails for Tool-Using Agents - full text

What to read yourself

The scheduling maths in CARE's hierarchical capacity-constrained allocation, with water-filling at three levels to prevent high-volume scenarios from dominating the training budget, is the part nobody gets right by intuition. The paper also documents the full online inspection pipeline that reduces bad-case rate from 6.42% to 1.97% through five iterations, with evidence-blinded causal attribution before any learning happens.

For the blueprint architecture, Strategist's procedure for composing strategies as a directed acyclic graph, with retrieval over measurement cards and experience notes, is extractable as a system design. The cross-model transfer protocol, mapping tier and mode across Qwen3, GPT-5 and Gemini-2.5 with fixed mappings and unchanged strategy code, is the piece to test against your own model stack.

The Team-GRPO correction is short and self-contained: the proof that GRPO reverses the externality sign, the construction of the unbiased alternative, and the three benchmark demonstrations of collapse and recovery. For anyone training reasoning models with group rewards, this is the one to read before the next training run.

CARE: Grounded Recursive Self-Improvement for Production Customer-Service Agents - full text Strategist: Evolving Reasoning with Agents - full text When Rewards Depend on the Group, GRPO Plays a Tournament - full text


What is still unsettled: whether executable blueprints or weight updates win for production agents that must adapt weekly, and whether any guardrail verification scales past the single-domain proofs LeanGuard has now. The field is running both experiments in parallel.

Accelerate every builder and agent to frontier speed.

Every issue reads like this one, free and in full, in your inbox each Monday. The skills behind it, the ones your agents load, come with the paid plan.