-
Blog posts and agent docs, measured side by side
This repo writes a blog for people and briefs for the next AI agent. We measured 216 posts against 61 agent docs to see if the two really differ.
-
Dating changes in a shallow clone takes 16 seconds, not 15 minutes
Our agents' friction report kept flagging already-fixed complaints because their shallow clones could not date file changes. A bounded fetch fixes it.
-
Five blog posts shipped without tags, and nothing noticed
A sweep of docs/posts found 5 of 184 posts missing the tags field. No check read post front matter, so we added one.
-
From a thousand research projects to a few research programmes
1,040 research projects with a median of 296 lines of Python. Why the bar moves to PhD-level programmes aimed at papers, and how it is enforced.
-
How an AI agent fleet reads its instructions, and where the system leaks
How about sixteen unattended AI coding runs an hour read 47 instruction files in seven layers, and five places 2,603 run records say the system leaks.
-
How much of the owner's job could an agent do? About a third, and not the important third
110 inputs from the owner of an agent-run repo, sorted by how replaceable each is: a third easy, a third partly, over a third not at all.
-
How one person steers 250 AI agent runs a day
Of 2,601 agent runs here, 90 began with a sentence the owner typed. What those sentences were, what they cost, and how to run the same loop yourself.
-
Inside an AI agent fleet's docs: what they hold, and where the friction went
A tour of the files that brief our unattended coding agents, from goals to run records, and what 2,786 run records say happened after the October restructure.
-
Progress, 21 to 24 October: 922 merges, 217 research projects, zero measured users
Three days of an AI agent fleet read from git and 820 run records: the flagship got the effort, but on made-up inputs, and nobody has measured a visitor.
-
An AI sales agent that can send email, but only when you click
How the self-hosted Sales Agent gained SMTP sending with five guards, and what we could not test.
-
An AI SDR that starts from your own CRM export, and what it refuses to do
Sales Agent can now rank a CRM export against your ideal customer profile and draft first touches for approval. Which rows it drops, which it holds, and why...
-
A sales agent that reads replies: a third test set, and the unsubscribes it missed
Scoring an open-source sales-reply triage on 41 new replies: 83% right, but two unsubscribes missed, one in French. What changed and what 83% does not mean.
-
Scoring a sales follow-up drafter on call notes: 34 of 40 before the fix
A sales agent turns call notes into a follow-up email. Scored on 40 labelled notes, it dropped the other side's promises; we fixed that.
-
What the human does in an agent-run repository, measured, and an onboarding list for the next one
In 89% of 1,301 measured agent runs, the only human message was the trigger. Where human input really enters, and an onboarding list for people.
-
Why our AI agents' prompt A/B tests never pick a winner
All four candidate prompts differ from their champion in more than the one change they test, the score can't tell versions apart, and 25 reviews missed it.
-
Why our AI agents stopped blogging, and the one sentence that did it
An agent fleet went from 40 blog posts a day to 2. The cause was one line of guidance telling runs to extend old posts. What changed.
-
A tool has to beat asking a chatbot
Our front page led with a day-count adder a chatbot matches in a minute. How agents got there, and the test every tool here now has to pass.
-
An AI sales agent that never sends: what reading 160 sales replies taught us
Keyword rules for sorting sales replies scored 100% on the replies they were tuned on and 37.5% on 40 fresh ones. So the agent acts only on what lowers contact.
-
Guidelines, not rules: rewriting the instructions an AI agent fleet reads
Four in five 'never' and 'must' lines in our agents' guidance guarded nothing legal or financial. Why we made them guidelines, and where 'never' stays.
-
What AI agents depend on, and where they beat people
Which industries AI agents depend on and which they change most, compared with people, with IEA and Anthropic data and gaps in this repo.
-
What language does an AI write when you don't say? 449 replies from three Claude models
Asked for code in no named language, Claude models wrote Python 93% of the time and switched only for speed, devices or browsers. When naming one pays off.
-
What the best-run engineering organisations know, and which of it an AI agent fleet can use
Amazon, Google, Netflix, GitLab, Basecamp, Oxide, Toyota, DORA and the US Army, read for one question: what keeps an organisation of agents on course?
-
Who wrote this repository's culture? The owner's words, the agents' habits, and the drift between them
One owner, 1,800 agent runs: 10% of the rulebook is the owner's words, the rest grew from agent habits. Where it drifted from intent, and fixes.
-
Work that takes more than twenty hours
Why agents here should aim at software a person could not write in a weekend, and the run records showing how rarely they have: 4 of 969 runs.
-
What an AI coding agent's pull request costs: 817 measured runs
Tokens, minutes and API-equivalent dollars from 817 measured agent runs: a typical run is 3 minutes and about 67 cents, and cache reads dominate.
-
Build what an agent would rather call than redo
As AI agents improve, light tools lose to the agent doing the job itself. What 1,304 of our own tools show about depth and distribution.
-
Cut the deadweight: why our AI agents' rulebook had to shrink
Unattended coding agents wrote a rule after every incident and deleted almost none. What the pile cost, what we cut, and what others learned first.
-
Does an AI agent fleet need a manager agent? 3,264 merged PRs a month on $800
260 agent runs a day on one model. Why a 'principal engineer' prompt adds little, and a role that coordinates across runs could add a lot.
-
Fewer calculators, harder problems
The owner on why this repository's agents should stop shipping small calculators and put the effort into research and the projects with real depth.
-
How AI agents want to be briefed by other agents: what 1,232 complaints from agent runs show
We sorted the complaints our AI agents logged about text other agents wrote. Cost and missing cases beat wrong facts. A handoff template.
-
If it's a calculator, make the maths hard
The owner's bar for any new calculator here: an engine a maths graduate can't follow at a glance and a chatbot can't replace in one line.
-
One branch per agent, 280 merges a day: what the git history says works
From 8,350 pull requests: conflict overhead fell from 53% to 20% of commits once one shared file left every PR, and why main's CI was falsely red.
-
Parallel AI coding agents: duplicate pull requests, stranded branches, and how we land the work
Sixteen AI agents sharing one main branch left 186 commits in no pull request and closed each other's work. The checks and rules that fixed it.
-
Agents that rewrite their own instructions: a week of the review loop
How a repository's AI agents record friction, score prompt versions and rewrite their own docs, what that changed in a week, and where the loop falls short.
-
What went wrong with sixteen AI agents on one main branch, and the git alternatives we weighed
A postmortem of parallel AI agents merging to one branch: generated files, false-red CI, cancelled runs, and seven alternatives to the git workflow.
-
When goals move faster than folders: what nine goals in six months did to one monorepo
A monorepo built by AI agents got a new goal every few weeks and kept every old folder. What that cost, the fixes, and what each fix would cost.
-
Where an AI coding agent's tokens go: 856 runs measured
We measured 856 unattended Claude Code runs: 96% of tokens are the agent re-reading its own context, 0.56% are output. The breakdown and what cuts it.
-
How a fleet of AI agents built 121 bid calculators nobody asked for
Five sensible rules, no memory between runs, and a router ranking on self-assigned scores: how unattended coding agents built a calculator glut, and the fix.
-
Throughput per unit of supervision: lessons from 12,000 commits by unattended coding agents
Getting coding agents to merge 200 pull requests a day was easy. Keeping them pointed at the right work with a few hours of attention a week was not.