Coding agents need traces, gates, and spend controls before autonomy is credible

Reliability work centered on who may act, what context an agent can see, and which tool calls need checks. The June 1 trend ties agent reliability to managed authority, diagnostics, and review paths. The June 6 trend makes the same issue product-facing: editable context, guarded desktop actions, and spending controls appear as practical requirements for agents doing real work. This makes control surfaces part of the agent design, not an afterthought added after deployment.

Training work used agent traces as more than logs. EvoTrainer lets diagnostics change as training branches fail or improve, using rollouts, configs, logs, and code diffs to decide whether to keep, prune, revert, or merge a branch. Socratic-SWE turns repository-solving traces into reusable repair skills, then generates targeted tasks checked by execution. The reported gains are concrete: EvoTrainer raises SWE-9B to 38.16 Avg@8 BC% against 33.77 for a human-engineered reinforcement learning setup, while Socratic-SWE reaches 50.40% on SWE-bench Verified after three iterations.

Evaluation work asked whether agents can handle the messy parts of software work: vague requirements, repository localization, tests, commits, and repeated feedback. Asuka-Bench scores web-app creation across multiple feedback rounds, with hidden requirements and browser-rendered behavior. TeleSWEBench builds 734 telecom tasks from real srsRAN 5G commits and reports that the strongest tools reach up to 25% ship-ready changes. These results keep autonomy claims tied to executable outcomes and domain code, where file localization and functional correctness both matter.

The week ended with a concrete production signal. The Economist excerpt reports Anthropic’s claim that Claude wrote more than four-fifths of the code it published in May, up from low single digits before Claude Code launched. The source gives no controlled safety result, but it does show the governance problem plainly: coding agents can now affect the systems, tools, and infrastructure that AI labs publish. Review, provenance, and limits on recursive software work need to be visible in ordinary engineering workflows.