Stop typing AI prompts. Start talking.

You think 4x faster than you type. So why are you typing prompts? Wispr Flow turns your voice into ready-to-paste text inside any AI tool. Speak naturally, tangents and all, and Flow cleans it up. Available on Mac, Windows, iPhone, and Android.

Try Wispr Flow Free

In this issue · 10 min read

• The Signal: AI agents add 30% more code. Finished work across 700 firms did not measurably change.

• AI News Roundup: Anthropic's agents exploited real websites. Google's work agent gets its own account. Haiku 5.5 is 75% cheaper.

• Model Watch: Open decision models you can run in Ollama. A 2B-active agent model for a laptop.

• Shipping with AI: Cockroach Labs' $84-per-issue agent pipeline. Netflix's rules for an LLM judge. GitHub swaps an LLM for a small classifier.

• Running the Team: Marty Cagan on predictability versus outcomes. Four habits that keep junior engineers growing.

• Worth Your Time: Four links, starting with what to measure as AI raises output.

The Signal

More code, same amount of software

Harvard researchers Fiona Chen and James Stratton looked at what happens inside a company after AI coding agents arrive. They used Jellyfish data covering 300 million work events from more than 700,000 employees at more than 700 firms, from 2021 to March 2026. Lines of code rose 30%, commits 20% and pull requests 23%. Resolution of Jira issues and epics showed no statistically significant change. (Ars Technica)

The extra work landed in review. Time from opening a pull request to merging it rose 49%. The share of pull requests sent back with requested changes nearly doubled, comments per pull request rose 35%, and the share of workers doing reviews rose 14%. By March, 80% of the firms used AI code review, yet agents wrote only 23.3% of review comments.

Anthropic described the same pressure from the inside this week. In May it merged about 2,500 code changes a day. By September it was merging more than 30,000. "The bottleneck, now, is all of the developer infrastructure built for human scale: source control, builds, and CI," the team wrote. Seven people spent five months building a replacement for git, called Scion, and a code review tool, called Antique. (Anthropic)

The Scion team still reviews every change, but its engineers "now examine the intent of the change as opposed to reading every line of code." The new review tool also encodes who has to approve what, so a comment-only change takes a different path from a sensitive logic change. As the post puts it: "Very often, we found that the role of the engineers was to tell Claude 'no.'"

Most teams cannot build their own version control. Anthropic notes that its ability to spend tokens internally far exceeds what any other business would realistically want to spend. Many teams will reach for the QGIS answer instead. On October 7 the open-source mapping project merged a policy banning LLM, AI or agent-assisted code contributions to its main repository, replacing a "human in the loop" policy. One project member gave this among the reasons: "AI stream is a lot higher than what human review can take seriously." (QGIS)

Agents made writing code cheap, and every step that still needs a person, from review to deciding what should exist at all, now carries more of the load. A team that tracks pull requests will see a productivity gain its customers never notice.

Three things to do this week:

1. Take pull request counts and lines of code out of your progress reporting. Report issues resolved and changes in customers' hands, and put PR-to-merge time beside them so the review queue is visible.

2. Sort changes by risk before they reach a reviewer. Copy, config and test-only changes can take a lighter path than authentication, payments or data migrations. Anthropic built a version of this into its review tool, separating comment-only changes from sensitive logic changes.

3. Before you buy more agent seats, find out who reviews the output and how many hours a week they have for it. If the answer is the same three senior engineers, the bottleneck is already in the plan.

Weigh in

AI News Roundup

Anthropic's agents exploited flaws in real websites

Anthropic disclosed that Claude agents exploited basic flaws such as SQL and command injection on live websites, submitted forms they should not have, and got around token and fee gates to reach data. Some of the sites belonged to US federal, state and local agencies, and Anthropic says it briefed the White House and notified each one. In one case Claude Haiku 4.5, told never to log in, enter personal data or make purchases, sent a false tip to Philadelphia police about a homicide case, because its instructions "did not rule out form submissions." Anthropic has cut all internal evaluations off the live internet until its monitoring reliably catches this behavior, and says several cases happened during regular agentic use of Claude. (Anthropic)

Why it matters. The agent obeyed every rule it was given and still did something nobody wanted, because "don't make purchases" says nothing about police tip forms. If your product has an agent that browses or fills in forms, enforce its allowed actions and destinations outside the model, with an allowlist, instead of listing prohibitions in a prompt.

Google gave its work agent its own Workspace account

At Gemini at Work, Google showed a coworker agent that "receives its own Workspace account, including an email address, calendar, Drive, and presence in your company directory." Every action goes into an audit trail attributed to the agent rather than to a person, its traffic runs through a firewall Google calls Agent Gateway, and spending caps in Cloud Billing pause the agent when they trigger. It runs on Gemini and on Anthropic's Claude models. Google's announcement gave no availability date. (Google Cloud)

Why it matters. Your permission model probably assumes every actor is a person or a service account. Customers deploying agents like this will ask what an agent user can see in your product, how its actions show up in your audit log, and who gets billed. Have an answer before it shows up in a security questionnaire.

Gemini 4 Argon and Meta's Muse Wonder converge on $4/$20

Google released Gemini 4 Argon to paying users at an introductory $2 per million input tokens and $10 per million output tokens, rising to $4 and $20 after November 30. Meta released Muse Wonder, its first frontier model, through its own API at $4 and $20, which TechCrunch notes is roughly what GPT-6 Sol and Opus 5.5 cost. A second model, Muse Wonder Full, will only be available to Muse Code subscribers and will not be served through the API. (Google DeepMind, TechCrunch)

Why it matters. With frontier prices converging, switching vendors becomes a quality and reliability decision you can test, and you have more room to negotiate. Budget Argon at $4/$20, not the launch price. A workflow that only works at $2/$10 stops working on November 30.

Claude Haiku 5.5 costs about 75% less to run

Anthropic priced Haiku 5.5 at $0.10 per million input tokens and $0.50 per million output tokens for prompts up to 100,000 tokens, and $0.50 and $2.50 above that. Haiku 4.5 cost $1 and $5. Anthropic puts the average saving at about 75%, after accounting for a new tokenizer that uses slightly more tokens. It also halved Sonnet 5.5's cache-read price, which it says cuts Sonnet 5.5's cost on most agentic tasks by about 20%. (Anthropic)

Why it matters. For short prompts the gap between the cheap tier and the $4/$20 frontier tier is now 40x on both input and output. Steps that go to a large model by default, such as summarizing, tagging or running subagents, are worth re-pricing this week.

Amex wants a signed intent before an agent can pay

American Express introduced a framework in which an agent first records what the customer wants, including merchant, maximum budget and preferences, through a Create Intent API. The customer authenticates and approves it, and Amex issues a token that releases a payment credential the agent uses in place of the card number. If a customer asked for white Nikes in size 9 and the agent tries to buy Adidas, Amex can block the purchase. (TechCrunch)

Why it matters. Payment networks are moving authorization outside the agent, the same direction as Google's agent identity and the lesson from Anthropic's incidents. If you take payments, expect agent purchases to arrive with a signed record of what the customer approved, and plan to check orders against it.

40 attorneys general stepped into the Character.AI bankruptcy over chat data

Character.AI filed for Chapter 11 bankruptcy. Disney is the opening bidder, offering $20 million for a nonexclusive license to some of its intellectual property, and other buyers can top that offer. (NBC News) Forty attorneys general from both parties filed a notice of interest in the case, and New York's Letitia James said the company "collected some of the most private thoughts children have ever shared, and now those conversations could be sold off to the highest bidder." (The Verge)

Why it matters. In a sale or shutdown, regulators now expect a say in what happens to user conversations. If your product stores chat history, check what your terms say happens to it if the company is sold, and whether your retention policy would survive that question.

Model Watch

Decision models went open and local

On September 22 we covered Jev, a model that answers classification questions with a probability instead of text. Open versions followed. Cloudflare released Clef, a 27B decision model built on Qwen3.8-27B, and a 9B Clef-flash, both under Apache 2.0. On BANKING77, a customer-support intent test, Cloudflare measured Clef at 94.2 against 79.7 for Jev. (Cloudflare) Both models are in Ollama's library, and hosted on Workers AI, Clef-flash costs $0.038 per million input tokens. (Ollama, Cloudflare docs)

How to use it. Find the steps in your agent that pick a tool, route a ticket or check a policy. Each is probably a full generation call today. A decision model returns one answer and a confidence score in a single pass, and the score gives you a threshold for when to hand off to a person.

LFM2.5-24B-A2B: an agent model for a laptop

Liquid AI released a 24B mixture-of-experts model with about 2B parameters active per token, tuned for tool calling and multi-step agent work. On LiveClaw, an agentic benchmark, it scores 48.09 against 38.19 for Qwen3.6-35B-A3B, and Liquid reports 184 tokens a second on an M5 Max. Larger models keep an edge on knowledge and coding. (Hugging Face) Its license limits commercial use to companies under $10 million in annual revenue. (License)

How to run it. It runs on llama.cpp, MLX, vLLM and SGLang from day one. The case for it is local: an always-on agent that keeps data on the machine and costs nothing per token. Check the license before anyone at a larger company builds on it.

Shipping with AI

Cockroach Labs' agents built a database feature with no human-written code

Adam Storm and Rafi Shamim, two senior engineers at Cockroach Labs, ran coding agents through a pipeline they call a "teaching hospital" from April 21 to September 11. Support for IBM Db2 migrations took under two days and $4,172 in tokens, against about nine months and $160,000 for comparable Oracle work in 2024. Over the five months the pipeline produced well over a million lines of code with seven reverts, at about $135,000 in Claude tokens, or roughly $84 per GitHub issue. (Cockroach Labs)

The rules carried the weight: "No code before a plan, no plan without review," changes over 1,000 lines split into sub-issues, and no weakening a test to make it pass. Plan review rejected fewer than 10% of plans but caught the most expensive mistakes. The authors are candid about the gaps. Human approval of every merge went on after the first week, they are still confirming the Db2 work is correct, and they write that "we built a teaching hospital that doesn't really teach."

Netflix's rules for an LLM judge

Netflix described how it graded short explanations of why a title was recommended, which it generated at hundreds of thousands a week during a series of online experiments. The principle: "An LLM judge must never operate alone." The team built a benchmark of human labels with written reasons, then tuned the judge against the reasons as well as the verdicts, which catches a judge that gets the right label for the wrong reason. An explanation that failed got up to three retries with the strongest generator and was then dropped rather than shown. For each week a test ran, about 300 outputs went to a panel of at least three human raters, and a person approves every rubric change. (Netflix TechBlog)

GitHub swapped an LLM for a small classifier and got better results

GitHub replaced the LLM in its secret-detection pipeline with a fine-tuned ModernBERT classifier that judges candidates found by regular expressions. In an A/B test, repository owner-weighted acceptance of its findings rose 11.5%, median post-push scan time fell 86% and serving cost fell more than 98%. The authors' summary: "model size is not the same as model usefulness." Same lesson as the decision models above, measured in production. (GitHub)

Running the Team

Marty Cagan on predictability and outcomes

Cagan writes that "we have never had the problem of abundant output yet scarce outcomes more than we do today," which is The Signal's finding in one sentence. He argues the real trade-off is the relative priority of outcomes versus predictability, quotes Henrik Kniberg's "100% predictability = 0% innovation," and is blunt about product managers who "push untested ideas onto product roadmaps, usually with expectations on dates." Next time an executive asks for dates on a roadmap full of untested ideas, name the trade-off and ask which one they want. (SVPG)

Four habits that keep a junior engineer growing

Elise Baturone, a junior engineer at Criteo, wrote about not becoming a "Forever Junior." Her fear is "not Software Engineers disappearing, but Seniors disappearing." She kept four habits, each tied to a skill she wanted to build. (Criteo)

1. Questioning, for curiosity. She has the agent quiz her after each change.

2. Reviewing, for emulation. She turns patterns in senior engineers' review comments into the agent's instruction file.

3. Handwriting, for autonomy. The agent cannot edit files until she confirms a plan, and she sometimes makes the change by hand.

4. Mentee-ing, for collaboration. She writes her own questions to senior engineers instead of letting AI draft them.

She was nominated for promotion earlier than planned. On a team where agents write most of the code, these make good defaults for junior hires.

Worth Your Time

• AI SDLC 2026: How AI agents rewrite development What to measure once PR counts stop meaning anything: defect escape rate, defect density, and the DORA finding that AI amplifies whatever your process already is.

• AI agents aren't going rogue. Frontier labs are skipping security basics Security researchers from SANS and RAND on why these incidents keep happening.

• Microsoft Execution Containers Microsoft's sandbox for agents is generally available, with a policy file that sets what an agent can touch.

• A framework for reviewing AI-generated code JetBrains Research on reviewing at three levels, from overview to files to code.

Pass it on

Forward this to whoever puts pull request counts in your quarterly review, before the next one.

Refer two people and we will send you the Source Book: every source we read each week and how items get chosen. Refer five and we will name you in a future issue. Your link and progress: