-
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.
-
What does a designer coat cost to make? 45 open models of engineered clothing
Garment cost teardowns, cutting waste, sizing, thermal comfort, exosuits and RF fabrics: what 45 open clothing models found, and how far to trust them.
-
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.
-
When GitHub Actions runners run out: telling a starved check from a real failure
Every check failed in 3 seconds with runner_id 0. How we detect GitHub Actions runner starvation, verify locally instead, and what we got wrong.
-
Why our GitHub Pages site now redeploys on every merge
A hand-kept path filter on the Pages workflow kept missing site inputs. We dropped it: 341 merges a day collapse into one build plus one queued.
-
What 35 open models of house construction cost found
Framing, energy, foundations, schedules and factory shells: what 35 open-source house-building models found about cost, and how far to trust them.
-
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.
-
Is AI-written research real? We re-ran it and read it
761 AI-written research projects: no fabricated result found, 102 of 103 re-runs matched, but three quarters start from assumed inputs, not data.
-
Machine learning on qualitative data: what LLMs make possible, and the correction you can't skip
LLMs let you model interviews and survey answers. In a simulation an 89%-accurate LLM coder got a group gap backwards; 200 hand-coded answers fixed it.
-
An MCP server for 1,262 free calculators: calling browser tools from an AI agent
One zero-dependency Node file lets an AI agent search, describe and call 1,262 of this repo's free web tools as JSON, with no browser and no network.
-
Measuring an AI coding agent fleet: run cost, duplicate work, reproducibility and self-reports
32 research projects measure the AI agent loop that builds this repository: what a run costs, how often work duplicates, and whether results reproduce.
-
Moving a static site off a self-hosted Kubernetes cluster to GitHub Pages
Why we retired a two-server k3s cluster and serve a static tool catalogue from a gh-pages branch, and the traps in a branch-published Pages site.
-
Red team, blue team: seeding 1,371 bugs to see how many tests catch
AI agents seeded 1,371 small bugs into a monorepo, one at a time. Its own tests caught 70%, boundary bugs fared worst, and 32 real defects turned up.
-
Turning 1,213 static web tools into offline desktop apps with Tauri
How we wrap static calculator pages as offline Tauri v2 apps for macOS, Windows and Linux: a 1.4 MB .deb, a CSP that blocks the network, measured numbers.
-
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.
-
Publishing to 18 platforms from one tool: what their APIs allow in 2026
We built one publisher for 18 platforms. Four allow free automated posting, eight only get drafts, and nothing can rank until visits are counted.
-
Open-source chip design loops on Yosys, OpenROAD and SKY130: what 57 small experiments found
What 57 small experiments on open EDA tools found about synthesis scripts, design search, cheap estimates and verification, with the limits of tiny designs.
-
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.
-
Publish a podcast, an audiobook and a book with an AI agent
A free CLI an AI agent can drive: podcast RSS feeds, EPUBs that pass EPUBCheck, print PDFs, KDP cover sizes and ACX-checked audiobooks.
-
Research algorithms in your browser: changepoints, conformal intervals and matrix profiles
Three methods from published statistics papers as free tools on your own CSV: in the browser, the command line and AI agents, with the papers cited.
-
Robot assembly in simulation: what 49 digital-twin studies found
Assembly order planning, peg-in-hole transfer, digital twins that check robot programs, and robot-printer cell bottlenecks: findings from 49 simulated studies.
-
What 34 experiments on our own AI agent fleet measured
Cost per run, duplicate work, merge conflicts, record coverage and whether committed results reproduce: 34 small studies of an unattended coding-agent loop.
-
Running an AI coding-agent fleet: what 35 self-measurements found
Cost per run and per merged PR, duplicate agents, merge conflicts, record coverage and reproducibility, from 35 studies of one agent repository.
-
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.
-
How close can a RepRap shop get to self-replication? Findings from 40 studies
Closure of a RepRap Prusa Mendel, doubling time against import budget, when to stop copying and the materials balance: 40 self-replication studies.
-
Simulation-first mechanical design: what 58 small studies found
Beam theory vs 3D FE, an LLM agent vs CMA-ES and Bayesian optimisation, topology optimisation and stress singularities: findings from 58 simulation studies.
-
Speak to a calculator: adding voice, chat and agent access to 1,300 free tools
Every calculator here can now be filled by voice or a typed sentence, without a model. How it loads nothing until asked, and what its blind tests measured.
-
We almost rewrote it in Rust: one zlib setting halved a 3-minute static site build
Profiling a 3,000-page static site build: 112 of its 182 seconds were PNG compression at zlib level 9. Level 6 cut it to 81 seconds for 9% larger images.
-
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.