Notes from the build.
Engineering deep-dives and design decisions behind Empryo.
engineering3 minCode has a shape. The bill knows.
One grep result cost us $1.14 — more than a rival's entire bug fix. So we rebuilt our engine, put it in a cage with the most minimal agent alive, and metered every dollar across six models. Pictures inside.
read →Aug 10
architecture11 minYour agent's memory is a search engine. Code memory should be a neighbor.
Empryo's agent writes its own memories and ties them to files. Touch a file, its memories arrive — badged with why. Rename the file, they fall away. Pin what matters; what gets ignored archives itself. Memory wired into the repo, not taped to the prompt.
read →Aug 2
engineering10 minI made my agent's prompt smarter before the expensive model ever read it
My first prompt compiler wrote hypotheses and dropped the fix rate from 3/3 to 1/3. The version that only points at files scored 3/3 at 11% less spend. Same model, same bug — the difference was deleting every field where a guess could live.
read →
architecture5 minMy code map was wrong about my own repo. So I taught it to learn.
A formatter in a side app ranked #1 across my entire codebase. Static analysis said it was important. My git history said otherwise. Here's what happened when I let the graph learn from how I actually work: 86 to 3,239 co-change pairs, and that formatter fell to #103.
read →Jul 17
guide8 minThe 10 best AI coding agents in 2026, ranked by someone who builds one
Ten agents ranked on what they actually cost and actually fix, measured against real provider invoices instead of vendor claims. Includes when to pick each one, and when not to pick mine (disclosure, I build #2).
read →Jul 16
benchmarks4 minI benchmarked two AI agents. Then I read the bill.
pi said $6.25. Anthropic billed $9.19. Empryo said $7.06 — billed $7.08. Self-reported agent costs are fiction until someone audits them. I keep auditing.
read →Jul 16
benchmarks4 minFive real bugs, two agents, and the one nobody could fix
We stopped writing benchmark fixtures and used real merged PRs from hono, zod and ky, too new for any model to have memorized. 7/10 vs 6/10, one 8-minute hang, and a bug that beat everyone.
read →
