Coding agents need factories, proofs, and controlled access

Several items describe agent automation as a repeatable line of work with explicit stop conditions. The practical recipe is narrow: pick a task class such as dependency updates, CVE fixes, flaky-test triage, or repo migrations; turn each job into a packet with scope, allowed tools, validation, no-op rules, and evidence; then stop at a named terminal state such as PR_READY, NO_OP, ESCALATE, or RETRYABLE_FAILURE.

The darker factory version uses the same core idea. Agents can run longer on maintenance work when tests, architectural decision records, or other checks judge the output. This keeps the safety claim outside the agent’s own explanation. A related piece names the missing pieces for larger teams: shared memory for requirements and decisions, plus controlled access to end-to-end tests and production-like environments.

Vericoding frames verification as the bottleneck after fast code generation. The proposed path starts with natural-language intent, translates it into Dafny-style preconditions and postconditions, checks the specification with Z3, generates code, and stores proof artifacts for audit.

The evidence is mixed in a useful way. The article cites a vericoding benchmark with 12,504 formal specifications and up to 82% success in Dafny using off-the-shelf large language models. It also cites gains in pure Dafny verification and AWS Cedar as production proof that formal methods can scale. The proposed end-to-end natural-language-to-verified-code product itself does not yet have a fresh quantitative evaluation in the item.

The tooling items treat repository hygiene as a direct cost for future agent work. Mcgoats packages an AI-assisted game repo with Claude Code instructions, continuous integration, branch protection, pull requests, auto-merge, post-merge tests, and test-driven-development conventions. Its practical contribution is the repo setup, not a benchmarked result.

ContextLevy targets a smaller but concrete failure mode: pull requests that add generated files, logs, snapshots, lockfile churn, or vendored code can inflate future agent context. It scans diffs, estimates context weight, and comments on risky pull requests without calling a model or uploading code. Together, these examples show guardrails moving into the repo itself: hooks, checks, permissions, and comments that shape what agents see and what they can merge.

The enterprise item centers on finance because the work has documents, policies, approvals, past outcomes, and human review points. Anthropic’s finance templates and OpenAI’s DeployCo are described as deployment programs with embedded engineers, business-process mapping, trusted data connections, backtesting, and review checkpoints.

The reported results are concrete but vendor-reported. OpenAI says its finance team processed five times as many contracts with the same headcount using Codex, and its internal IR-GPT handled more than 200 investor interactions. Anthropic reports Claude Opus 4.7 at 64.37% on Vals AI’s Finance Agent benchmark. PwC cites an insurance underwriting cycle cut from 10 weeks to 10 days with backtesting and human oversight.