Coding-agent research is making structure and external checks the main control surface

Coding-agent papers for this period put more weight on explicit, checkable control. One proposal frames pair programming around a driver and a navigator, with the navigator producing machine-checkable contracts and formal specs, then validating code and tests with deterministic verifiers and SMT-backed counterexamples. The end-to-end workflow is still a research plan, but its supporting systems report concrete verification results: AutoReSpec verified 67 of 72 programs, and AutoJML verified 109 of 120.

A separate repository study finds that giving agents an architecture descriptor cuts code-navigation steps more than it lifts raw task accuracy. On a 22K-line Rust project, architecture context reduced average navigation steps from 5.2 to 3.4 with S-expressions or JSON, and to 2.9 with Markdown. In a second study, an auto-generated 170-line descriptor reached 100% accuracy on 15 tasks versus 80% for blind search. The common point is practical: agents do better when key structure is written down in forms that tools can reuse and check.

The strongest infrastructure result in the set comes from memory management inside the agent harness. ClawVM treats agent state as typed pages with minimum-fidelity rules, then assembles prompts in two stages: keep the required state first, spend remaining tokens on higher-value detail second. That is a concrete answer to a familiar failure mode in long sessions, where plans, constraints, or prior evidence disappear after compaction or reset.

The reported gains are large. Across four workload families and six token budgets, ClawVM reduces mean policy-controllable faults from 67.8 in a retrieval baseline and 1.5 in a compaction-plus-retrieval baseline to zero when the minimum required state fits. On 12 real traces and 30 task replays, it reports the same zero-fault outcome and 100% success at the tightest budget, with median policy overhead under 50 microseconds per turn. That keeps memory handling in the harness, not in fragile prompt text.

Benchmark work is getting closer to real operating limits. USACOArena charges agents for tokens, tests, and elapsed time, then ranks them with ICPC-style scoring. That setup exposes a blunt fact: frontier coding agents still leave a lot of room on the table when cost matters. Across four contests, the theoretical maximum score is 54 points, while top agents average around 15.

Gemini-2.5-pro and GPT-5-Codex lead the compute-only setting, but the benchmark is far from solved. Gemini posts an average rank of 1.3 with a 70% win rate, and on the USACO 2025 US Open it scores 14.6 against 3.0 for GPT-5-Codex. Budget scaling is also uneven. Cutting Gemini's credit limit to 10M drops its score from 13.2 to 8.3, while raising the budget to 40M leaves it near 13.0. More spending does not reliably buy better play.

One paper pushes coding agents into direct experimental work on released research code, and one systems paper applies LLMs to database execution plans. The algorithm-improvement study reports gains in all 11 selected implementations within one working day, including 193x runtime improvement in combinatorial optimization and more than 1000x faster image segmentation at high K. The paper also states that these results were not independently rerun outside the agent environment, so the headline numbers need caution.

The database paper is more narrowly scoped and more concrete about mechanism. It has the model edit existing Apache DataFusion physical plans through JSON Patch operations, then keeps only rewrites that improve execution. In one TPC-DS-style case, moving a year filter earlier reduced a sales table from 15.1 million rows to 2.9 million before later joins, producing a 4.78x speedup and cutting hash-table build memory from 3.3 GB to 411 MB. This period includes several agent papers, but the applied systems result is one of the clearest user-visible wins.