Engineering Notes on Agent-First Software
How AltSlate Labs builds agent-first software and the compact models behind it: what we tried, what we measured, and what did not work.
All articles
An AI That Grades Its Own Code Keeps Its Own Misreadings
We trained a model on answers from running its own code. A misreading in that code never got fixed. What did fix it, and what that fix then broke.
Fine at 32 Bits, Broken at 4: Where the Softmax Bug Shows Up
We mapped where a known training bug shows up. In 32-bit training it barely matters. In 8- and 4-bit it hits most of every batch, attention first.
A New Fix for a Training Bug Is Real. The Accuracy Gain Mostly Isn't.
We re-ran a softmax loss fix in PyTorch across 480 training runs. The bug is real, the accuracy gain small. Where 16-bit maths happens mattered more.
A 16‑Line Code Check Beat Automated Prompt Tuning
Every rewritten prompt made our small-model chatbot worse, automatic or not. A 16-line check that refuses incomplete records fixed it. Five lessons.
RL Does Exactly What You Reward. The Environment Decides What That Is.
We built an RL environment for oil-well troubleshooting. RL optimised exactly what we rewarded, loopholes included. The worked examples did the rest. Four lessons.
AI Confidence Scores Depend on the Options You List
Adding an option, even one nobody picks, moved a decision model's scores. Longer lists moved them more. What we measured and five practical lessons.
Fewer, Closer Examples Beat a Smarter Lookup
We tried making AI lookups push look-alike examples away. It failed. Reading fewer, closer examples worked better. Three practical lessons for noisy data.
Image Checks From a Few Dozen Labelled Photos
Calibrated yes/no, choice and score answers about images from a few dozen labelled photos, with an abstain flag. Coarse checks hit 93–95%; fine detail fails.
Speed or Rules? What Caching a Decision Model Costs
Storing options made our decision model 5× faster on big menus but worse at stated rules. Training restored it on synthetic rules; real ones are unproven.
Agent Harness: How We Run 40 AI Agents in Production
What an agent harness is and how ours runs about forty AI agents in production: one graph, permission gates, typed tools, human approvals and evals.
The Assistant That Asks Before It Acts
An operations assistant where AI agents do the legwork and changes to the business arrive as previews a person approves. What it is like to use.
Half the Memory, None of the Answers Lost
Small AI models get expensive when they must remember a long history. One fix flopped. The other halved that memory for free, without changing an answer.
The Verifier's Blind Spot
We built a "self-improving" LLM cascade to get cheaper every round. It got quietly worse — and every dashboard said it was fine. What we measured, and why you cannot see it from inside.
Technology Finally Reached the Delivery Layer
Cloud rebuilt the front door of services firms and left the workshop untouched. Agents walk into the workshop — and that changes the economics of professional services.
Why We Build Agent‑First, Not Agent‑Bolted‑On
Bolting an LLM onto an existing app gets you a demo. Designing the agent as the operating core — with humans in command — gets you a product. Here is how we think about the difference.
Serverless That Actually Scales (and Stays Cheap)
Serverless is not automatically cheap or scalable — it is a set of trade-offs. Here are the patterns we lean on to get fan-out compute that bills to zero when idle.
Production, Not Pilots — How We Build
Most AI work stalls at the demo. We design, build, and operate production systems end to end — applying AI where it earns its place and conventional engineering everywhere else.