Engineering field notes

Your agent is not slow because the model is weak. It is slow because every session it re-derives what the repository is, and nothing it learned yesterday survived. Here is where the tokens actually go, why a bigger context window does not fix it, and the three layers that do.

Running one coding agent is a tooling problem. Running five on the same repository is a distributed-systems problem — double claims, silent hangs, orphaned children, lost output. Four failure modes, and the minimal protocol that removes them.

An agent that starts every session from zero re-asks the same questions and repeats mistakes it already paid for. A memory store fixes that only if it is designed against the real failure mode — which is not 'nothing found', it is 'confidently wrong'.

Most retrieval layers return the top ten matches and let the model pay for whatever they weigh. A budget-aware retriever answers a different question: return exactly enough material to answer X within N tokens. Here is how that is built, and how to measure whether it worked.

Most teams keep architecture decisions on a shelf nobody rereads, so agents keep proposing what was rejected six months ago. Turning those records into a validated graph — with stable identifiers, one edge kind, and a build that fails on a dangling reference — makes the reasoning machine-readable.

A human reads an answer and senses when something is off. An agent acts on it immediately, unsupervised, and cannot see what produced it. That one difference rewrites what a service owes its caller — starting with the rule that an answer is a claim about its own reliability.

Comprehensive comparison of 10 leading Generative Engine Optimization (GEO) and AI visibility tracking tools in 2026 — Profound, Athena HQ, Otterly, BrandLight, Peec, Semrush AIO, Ahrefs Brand Radar, and more. Pricing, features, who they fit.

I ran a 26-criterion AI-readiness audit on ivannikov.pro. Score: 42 → 80 in one day. Then I ran the same audit on Stripe (55), Twilio (UNRELIABLE), and Linear (UNRELIABLE). Here's the methodology and what it means for B2B SaaS in 2026.

A comprehensive analysis of the current state of Web3, DeFi, and the blockchain market.