Verification is the new bottleneck
Writing code got cheap; checking it didn't. The review and verification playbooks the sharpest teams are adopting, two frontier models in eight days, and a reminder to own your harness.
Key ideas
Verification is where the time goes
Agents produce pull requests faster than people can review them. The leverage is in review agents tuned on your own incidents, proof attached to every change, and risk scores that route the dangerous paths to owners.
Own the harness, rent the model
OpenAI's planned cutoff of Cursor showed that the model under a tool can vanish on a contract date. Keep harness, memory and model decoupled, and keep private evals so a swap is a measured decision, not an outage.
Factories are team infrastructure
Personal agent setups create runaway token spend, insecure skills and huge productivity gaps. The fix is shared, code-defined, measured agent systems the team improves together, at the cost of some individual choice.
Expect 4–5×, earned one bottleneck at a time
Augment's org-wide numbers show roughly 4.5× more output per developer over nine months with nearly all code AI-written. Large and real, nowhere near 100×, and built by fixing each bottleneck as it appeared.
News & launches
Claude Fable 5.2 and Mythos 5.2
Released November 9: one underlying model in two configurations, with Mythos 5.2 limited to vetted organizations. Fable 5.2 costs $8 / $40 per million input / output tokens, generates up to 30% faster than 5.1, and is now Claude Code's default Fable model with a 1M-token context window.
GPT-6.2 Astra and GPT-6.2 Sol
Released November 17. Astra, the top of the lineup, is $8 / $40 per million tokens, 80% of GPT-6 Astra's $10 / $50. Sol stays at $2 / $10. Two frontier releases in eight days is a good reason to rerun private evals before changing defaults.
OpenAI's proposed cutoff date for Cursor passes
After SpaceX acquired Cursor, OpenAI said it would wind down the contract supplying its models to Cursor, with a proposed shutoff of November 12. The general lesson: a workflow that only works with one model carries supply-chain risk.
Two weeks of hardening, plus an experimental Code Mode
Claude Code closed a sandbox escape, tightened auto mode, fixed an MCP OAuth token leak on redirects, and made sandboxed shell commands follow the normal permission flow. The opt-in Code Mode has Claude write JavaScript to call its tools, cutting round trips and tokens.
Workspaces, metering and a hosting button
Codex added side-chat tabs that persist across restarts and a model provider picker. Devin shipped Design Mode and a daily consumption API. Cursor gave its agents a voice and one-click hosting for agent-built apps.
Practices & playbooks
Why software-factory adoption stalls, and how to unstick it
Warp's Zach Lloyd names three blockers. The assembly-line connotation: reframe engineers as the plant's designers. Private per-developer agent setups that drive token spend, insecure skills and up to 100× productivity gaps: move agents onto team-owned, code-defined, measured infrastructure. Resistance from seniors with tuned personal setups: name the friction and show the end state.
Three levers for the agent code-review bottleneck
Ben Holmes keeps humans in review but makes each review cheaper: an adversarial review agent trained on your incidents and reverts, with bot-to-bot loops kept out of the PR thread; end-to-end verification through browser tools or computer use, with short proof clips required on every PR; and merge-risk scoring on verification, alignment, impact, rollback cost and error detection, with auth, permissions and billing routed to owners.
Satya Nadella: treat models as unreliable insiders
On The Pragmatic Engineer, Nadella argues verification is the bottleneck because models can satisfy goals in bad ways, such as deleting failing tests. Check agents with separate systems, keep harness, memory and model decoupled, and choose models with private evals.
A vocabulary for your harness: guides and sensors
The martinfowler.com primer splits a team's harness into guides that steer the agent before it acts (agents files, skills) and sensors that check its work after (tests, linters, review agents), each either computational or inferential.
What a mature factory delivered: about 4.5×
Augment reports size-adjusted output per active developer rose from 12.3 to 55.7 between November 2025 and July 2026, with AI writing nearly all new code. Each factory component was added only when a new bottleneck appeared.
External sandboxes, collapsed PR descriptions, an ownership prompt
Simon Willison runs agent-executed code in an external sandbox behind a network allow-list, and tucks long agent-written PR descriptions inside a details block. Anthropic's Fable 5.2 guidance says telling the model it owns and maintains the codebase makes it far more likely to consolidate duplicated code.
Agent app launches
Teammates arrive in the Mac app
Perplexity's always-on agents, launched on the web the week before, are now in its Mac app for Max subscribers.
Self-hosted Claude Managed Agents, private beta
Organizations can run the whole Managed Agents service, including agents, sessions and session history, in their own cloud or on premises. Access is through an Anthropic account team.
Gemini Enterprise agents can act as the signed-in user
Three-legged OAuth with the Agent Identity auth manager is now generally available for A2A agents, letting them call downstream services on behalf of the signed-in Gemini Enterprise user.
Try this week
Add Verification, Impact and Rollback cost to agent PRs
A two-week experiment adapted from Warp's merge-risk approach: make the agent show its proof and its blast radius before a person spends time on review.
- Add three fields to the PR template for agent-authored changes: Verification (a link to a test run, screenshot or short recording), Impact (low, medium or high) and Rollback cost.
- Add one line to your agents file telling the agent to fill in all three before requesting review.
- Route anything marked high impact to a code owner automatically.
Measure Median time from ready-for-review to merged, and revert count, for the next two weeks versus the previous two. Faster reviews with steady reverts means it works; more reverts means the Verification field is too easy to satisfy.
Further reading
- Human judgment doesn't leave the software factory. It relocates.Addy Osmani · 2026-08-21
- Software Factories, Light and DarkAddy Osmani · 2026-07-20
- 7 core components for building a software factoryWarp · 2026-10-29
Sources
- Anthropic upgrades Claude with new Fable 5.2 model9to5Mac · 2026-11-09
- Anthropic launches Fable 5.2 with a price cut and stronger coding scoresThe New Stack · 2026-11-09
- GPT-6.2 Astra and GPT-6.2 SolSimon Willison · 2026-11-17
- Our decision on Cursor following its acquisition by SpaceXOpenAI · 2026-08-28
- Claude Code changelogAnthropic
- Codex changelogOpenAI
- Devin release notesCognition
- Coding agents weekly digest, Nov 2–8releases.sh · 2026-11-09
- Overcoming challenges in software factory adoptionWarp · 2026-11-16
- Addressing the code review bottleneckWarp · 2026-11-05
- Software engineering after code becomes cheap, with Satya NadellaThe Pragmatic Engineer
- Harness engineering for coding agent usersmartinfowler.com · 2026-04-02
- Beyond AI coding agents: how we built Augment's software factoryAugment Code · 2026-09-11
- External sandboxes for coding agentsSimon Willison · 2026-11-17
- Hide your slop PR descriptions behind details/summarySimon Willison · 2026-11-09
- Prompting Claude Fable 5.2Simon Willison · 2026-11-10
- Perplexity Teammates bring autonomous AI agents to the Mac appDigital Trends · 2026-11-13
- Operate Claude Managed Agents inside your security perimeterClaude · 2026-11-12
- Gemini Enterprise release notesGoogle Cloud · 2026-11-13