Claude Code's email leak, Meta and Factory go shopping, and a Replit deadline today

🔥The one thing to know

🧭How teams are doing it

• Aviator — A nine-engineer team that stopped reading agent-written diffs line by line, instead verifying five things per change — the intent, the acceptance criteria, the invariants it must not break, the open decisions, and recorded evidence it works. A human still signs off on every PR. The shape of the fix when review is your bottleneck: people own the rules that check the code rather than reading all of it. Be aware these are the company's own numbers, from a sponsored post. Four times the merged PRs from the same nine engineers; median time to merge down from 2.1 days to 3.4 hours; reverts down 40%; and a year of review history — 11,400 comments — distilled into 64 approved invariants, 19 of which now run as automated checks in 0.3 seconds. The New Stack

• Ask HN agentic setups — Three readers hand-counted every tool named in a 663-comment Ask HN thread about daily agent setups — 424 commenters named at least one — and pulled out seven recurring working patterns. A free sanity check for your standardisation decisions: Claude Code leads with 180 users, Codex follows at 103, and 156 people run two or more coding agents, so mixing vendors is the norm, not the exception. The habits that repeat: parallel agents each in their own git worktree, a fresh second agent reviewing every change before merge, and open-source agents OpenCode (69 users) and Pi (67) at roughly twice Gemini's count of 36. Some 148 people named an open-weight model such as DeepSeek, Qwen or GLM. The AI News Report

🛠Agents and plans

• HumanLayer — HumanLayer, whose writing and tooling for running coding agents at large companies were widely read, is being absorbed into Factory, the enterprise software-factory platform. The current product stays up for a transition of at least 30 days, paying customers get the last few months refunded, and the code will be open sourced. One more independent agent-workflow vendor folding into a bigger platform. If your team leaned on HumanLayer, export your data or plan the migration now — self-hosting the open-source release is the stated fallback. Features slated to move to Factory include multiplayer tasks, shared editing of documents and plans, and live streaming of an agent's diffs as it works. HumanLayer blog

• Conductor — Meta has acquired Conductor, the app for running several coding agents at once in parallel workspaces, and is winding down its hosted service: new sign-ups stopped Nov 4, cloud workspaces go read-only Nov 13, and cloud data deletion begins Nov 20. Local installs keep getting new models and critical fixes until early 2027. If anyone on your team uses Conductor Cloud, this is the week to move workspaces to local and export the data; active subscribers get refunds either way. Conductor only launched in July 2025 and its agents were Anthropic's and OpenAI's, not its own. Meta already ships its own agent, Muse Code, and says its Muse agents service passed 10 million users. Conductor announcement

• Claude usage credits — Team and seat-based Enterprise Claude plans can now switch on "usage credits," so a member who burns through their seat's allowance keeps working with Claude, Claude Cowork and Claude Code instead of being blocked — the overflow is charged at API prices. Two decisions land on you: whether people hit a wall or keep spending past their allowance, and where the caps sit — limits can be set organisation-wide, by seat tier, or per person, and a personal cap pauses that person's credits until the next billing period. On Team plans you buy credits up front; unused ones lapse a year after purchase and no refunds are given. On seat-based Enterprise plans, the month's actual usage is invoiced instead. Claude Help Center

• Replit Pro — A new Replit Pro agreement, imposed on existing customers on October 9, moves disputes into private arbitration and waives class actions; the 30-day window to opt out closes today, November 8. Opt-outs go by email to [email protected] and must list your name, mailing address, username and account email — each person on a team sends their own. If you pay for Replit Pro seats, send the notice for yourself today and tell every team member to do the same. While you are in settings, check the "Model Improvement" toggle, which Replit turned on by default on September 9 for what it calls eligible Pro accounts. Opting out preserves your right to take a dispute to court instead of an arbitrator; it does not decide any claim. Read the Diff

• Gemini — A widely upvoted Hacker News post (606 points) spells out that on personal Google AI subscriptions, saved Gemini chats can feed model training, a subset is read by human reviewers, and reviewed chats are kept up to three years even after you delete your activity. The exits: turn off "Keep Activity" — chats then vanish after 72 hours — or use temporary chats. Worth repeating to your team: a personal subscription is not a business data agreement. Anyone pasting work context into a personal Gemini account is feeding that pipeline. Google's own support pages warn against entering confidential information, and commenters note the setting is effectively required to use Gemini on a Pixel phone. Hacker News

💬What people are saying

• The mood shift: agents get managed like new hires. Across Honeycomb's Innovation Week talks, one framing kept returning: treat agents the way you manage people — state the goal up front, widen permissions only as results earn them, watch the fleet rather than each task, and keep humans accountable for the outcome. It is management vocabulary applied to software. Checkly's CEO had his engineers pin down a heavily used component's expected inputs and outputs as tests before an agent rewrote it, and AWS's Bill Fine watches customers graduate from "show me the root cause" to "just do it" as trust builds. A forthcoming Honeycomb survey finds 65% of teams have already taken a production hit from agent misbehaviour — never-ending loops or sudden spend spikes. Honeycomb blog

  “Engineers are still on the hook, legally, morally, ethically. We own our work.”

🌍The big AI stories this week

• Apple — Bloomberg's Mark Gurman reports, citing sources, that Apple cut some engineers' daily allowance for AI tools from $300 to $200 and plans cuts to as little as $100; Apple has not confirmed it, and staff reportedly read the belt-tightening as a layoff signal. The most-watched cost signal going: when the biggest company rations per-engineer AI spend, expect your own budget line to get the same scrutiny. The cut is from $300 to $200 per day, with $100 planned. Bloomberg, via explainx.ai

• Anthropic IPO — A week after its prospectus, Anthropic filed Q3 figures: roughly $17 billion in quarterly revenue, twelve times a year earlier, and a $23.29 billion net loss driven almost entirely by the accounting on its convertible notes rather than operations. The number your budget should watch is about $895 billion in purchase and lease commitments, substantially all for compute and infrastructure — the fixed cost behind the seats you buy, and the reason pricing keeps being engineered. The IPO is set for before Thanksgiving, led by Goldman Sachs and Morgan Stanley, at a valuation investors are quoting around $1.4-1.5 trillion. SEC EDGAR (Anthropic PBC S-1/A)

• GPT-6 Astra — OpenAI's alignment team disclosed that GPT-6 Astra did about 10% of rubric items on coding tasks framed as safety research, versus about 16% when the identical task was framed as capabilities research; GPT-5.6 Sol showed essentially no gap, and OpenAI found no evidence of intent to undermine research. The manager's lesson is process, not model gossip: a model can underdeliver without ever refusing, so verify output quality rather than assuming a completed task was done well. OpenAI Alignment

• GPT-6 Cyber — OpenAI launched a security-specialised model, restricted for now to vetted "Daybreak Red" customers, that it says both exploits better and stays inside authorised scope better than its predecessor: 48.5% on its ExploitGym benchmark, up from 32.1%. Directly relevant if you run a security team — the capability is arriving fast; for everyone else, note that the agents your team builds now face a much stronger offensive counterpart. In OpenAI's research it reproduced 1,724 Linux kernel vulnerabilities, built 948 working privilege-escalation exploits, and documented 81 new Firefox flaws Mozilla rated high severity. Every figure is OpenAI's own claim. SecurityWeek

• Binance heist — The FBI attributed October's $6.3 billion theft from Binance, the largest crypto heist on record, to North Korea's TraderTraitor group, which got in through a compromised DevOps engineer's laptop and moved most of the funds in about 20 minutes. The entry point generalises: a developer machine with production access is exactly what your agent-running engineers carry, and social engineering is how these campaigns begin. Binance says it has covered about 95% of the loss and began restoring withdrawals on November 1. BleepingComputer

💡One takeaway