Engineering deep-dives, benchmarks and customer stories on persistent code context, knowledge graphs, and the AI coding tools we work with daily.
21–30 of 111 · page 3 of 12
Compliance with an agent's own explicit constraints falls from 73% at turn 5 to 33% at turn 16 across 4,416 measured trials, and Claude Code auto compaction reduces 132,000 tokens of session state to roughly 2,300. Shorter sessions and cleaner instruction files treat the symptom. The window fills with low value tokens because the agent rebuilds an understanding of your codebase from raw source on every task.
DeepSeek published its agent runtime on 13 August 2026 and it passed a hundred thousand stars in two days, but almost everything written about it was the install command. This is a plain English read of the source. What an agent runtime actually is, what everything is a plugin means in practice, the three layers inside every capability, the four modes, the security defaults, what the benchmark numbers really measure, and the rough edges nobody is listing.
This argument was being compared to the Vim and Emacs wars in 2024, and then it returned everywhere inside six months. Something dragged it out of retirement, and it is not really about where files live. The variable that decides whether agents work is whether the question of what depends on this is answerable at all, and neither layout answers it.
We pointed DeepSeek Harness at an MCP server we already ran and it worked without a line of new code. This is what the official bridge carries, what it quietly drops, where plugins get published, and the one reason a verification tool cannot live behind MCP at all: over MCP the model decides whether to call you, and a check the model can skip is not a check.
Agent runtimes got very good very fast. The thing that decides whether a change is correct across your other repositories did not get built at all. On cross repository blast radius, why AI code review misses it, and what a check actually needs.
A technical note on what 'verifiable' actually means in a verifiable code IR. Not a machine that rebuilds your code from a document, but a derived contract that every change gets checked against. Here is what a single verification check consists of, why round-trip is a signal and never a promise, and the honest boundary of what it can and cannot catch.
A code graph takes a whole, coherent program and shreds it into a bag of symbols and edges. It is genuinely useful, but it is a teardown, and the meaning never survives the pieces. A verifiable intermediate representation is the opposite move. It lifts your code into a higher layer that keeps the intent intact, then checks every change against it. Here is the difference between shredding and preserving, and why it decides everything downstream.
GitHub Copilot moved to per-token billing and one developer's bill jumped from about $29 to a projected $750. Switching tools won't save you, because the same shift is happening everywhere. Here's why your bill exploded in plain English, the practical levers that cut it today, and how verifiable specs attack the cost at its root.
Per-token billing turned a flat $29 AI coding plan into a $750 surprise. The waste was always there: most of the bill is your AI re-reading code it already saw. Here's why it happens and how to cut it, with prompt caching, the right model, the right plan, and a lasting memory of your codebase.
Vector search and GraphRAG look like opposites, but they share the same final move. Parse the code, retrieve some of it, rerank it, and stuff it into the context window. That last step is where both of them rot, because a model that is handed a pile of chunks degrades the same way no matter how the pile was chosen. Here is why the retrieval layer is not the cure for context rot, and what is.
