Coding agents need census data, cost controls, and security evidence

The clearest empirical signal is the multi-method census of AI coding agents across more than 180 million Git repositories. It shows that bot-account lookup alone misses most activity. For Claude Code in the V2510 snapshot, the multi-method scan finds 850,157 commits, while bot-account lookup finds 28,154. The V2604 snapshot attributes 1,772,677 commits to agents, with Claude Code accounting for half. This gives supply-chain studies a stronger base than single-channel traces such as pull requests, bot names, or configuration files alone.

Agent work is being treated as a control problem with explicit costs. The Bayesian control paper keeps a posterior belief that a candidate program will pass verification, then decides whether to run cheap critics, regenerate, verify, or stop. Its evaluation spans six generators and nine coding benchmarks. The reported gains are strongest when verification is expensive and cheaper critics give imperfect signals. BigBag applies a related discipline to dependency repair: it asks an agent to generate executable Java abstract syntax tree transformations, then reuses those transformations across clients affected by the same library update. The best setup reaches a 78.6% fix rate on original broken clients and a 33.3% cross-project fix rate overall.

Security work in the corpus asks for executable proof and readable controls. Kauge separates secure-coding knowledge, code actuation, and the gap between them, using OWASP and CERT principles plus exploit tests. The paper reports that models often know the relevant principle yet fail to put the defense at the right boundary. AutoSpec addresses a different part of the safety problem. It revises symbolic safety rules for large language model (LLM) agents using labeled execution traces, then keeps edits that improve trace-level scores. On 291 traces across code-execution and embodied-agent domains, it reports F1 scores of 0.98 and 0.93 and up to 94% false-positive reduction. Postman Passport adds a product-side example: API callers receive credential references, while real keys stay in the customer vault and are resolved through a proxy with scope checks.

Several items apply multi-agent systems to research and software knowledge work. Agon runs reusable producer-critic loops for ideas, proposals, experiments, and papers. Its evidence is operational scale: 444 loop iterations, 18 roles, projects across more than 10 scientific domains, and a month of dispatcher operation. The paper is careful that claim judgment still needs humans. LLM4MTLs adds a smaller, more testable software-engineering case. It evaluates generated model transformation language code across 47 examples, four languages, and three large language models. Few-shot examples improve syntax across all four languages, while semantic correctness depends on the language and task.