CONTEXT ENGINEERING

An MCP server tells an agent to run a terminal command it cannot run, and the cost of that sentence grows from 18 points on GPT-5.5 to 69 on GPT-6 Astra. A pipeline loses up to 40.5 points at its own interfaces, and one instruction clause swings 30. Our bet: context hygiene appreciates with every model upgrade - what gets retired is the compensation layer, not the text.

Agent Retrieval Bench isolates the part of coding-agent work most evals hide: did the agent find the right repository files before it started editing?

CodeNib's July paper argues that coding agents should stop rediscovering the same repo through grep and reads. Repository context is becoming compiled infrastructure.

The MCP maintainers published an updated roadmap on August 22, 2026 with five priority areas, including progressive discovery for tool catalogs and standardized agent identity. Here is what changes for developers building MCP servers and agent platforms.

A study of 247,694 instruction lifetimes in 1,867 repositories shows agentic prompt files grow +226% on average because the reasoning behind each rule decays. Comments encoding that reasoning remove 99.3% of the excess.

A late-July research wave - native in-backbone memory, pretrained parametric memory at scale, memory reconstruction, and transactional memory writes - challenges the external-store paradigm every agent memory product is built on. Here is what changes by late 2027 and what developers should do now.

A SCAM 2026 study of 100 top-starred repos catalogs six configuration smells in AGENTS.md and CLAUDE.md files: Lint Leakage in 62%, Context Bloat in 42%, Skill Leakage in 35%. Only 9 of 100 files were smell-free.

A controlled ablation across Claude Code and Codex, 17 real tasks, and 288 evaluated runs finds context-injection strategy does not measurably change correctness (bounded to under 10-15pp). The failures are implementation skill, not missing repository knowledge.

A new paper wraps code into an OWL2 ontology with SPARQL property paths to answer multi-hop structural queries for coding agents - 2.06x retrieval recall and 28.8% fewer tokens on SWE-bench Pro, versus treating code as plain text.

The Agent Skills spec gave agents progressive disclosure in three tiers - name, SKILL.md, bundled files. What it did not give them is a graph. Skills that link to each other, and say when to follow the link, let an agent navigate knowledge instead of front-loading it. Here is the argument, the measurements from our own 36-skill repo, and what to change.

Anthropic cut 80% of Claude Code's system prompt for Opus 5 and Fable 5 with zero regression on coding evals. The post landed on HN with 197 points and 133 comments. Here is what the article says, what HN thinks, and what it means for your agent harness.

Anthropic removed over 80% of Claude Code's system prompt for Claude 5 models. Here is how the rules changed and what it means for your CLAUDE.md files, skills, and system prompts.

A deep dive into why fully autonomous AI coding agents degrade codebases over time, and what context engineering can actually fix.

SWE-Pruner Pro points at a practical coding-agent design shift: do not only compress prompts outside the model. Teach the runtime to prune tool outputs before they become the next turn's context.

Lilian Weng argues self-improving AI won't start with models rewriting their weights - it starts with the harness. Here's what that means for developers building agents.

Fable 5 1M context workflows that actually work: whole-repo reviews, log archaeology, multi-doc synthesis - plus the honest math on when RAG still wins.

GitHub Trending is full of agent memory and context tools. The useful version is not magic recall. It is a context ledger: source-linked, scoped, expiring memory that agents can inspect and users can audit.

A huge Hacker News thread says domain expertise is the real moat in agentic coding. The sharper version: tacit judgment only compounds when you turn it into examples, tests, DSLs, and review gates.

GitHub is suddenly full of codebase knowledge graph projects for Claude Code, Codex, Cursor, and other agents. The useful version is not a pretty graph. It is a map that changes planning, editing, and review.

Persistent memory for coding agents is trending because every session still starts too cold. The hard part is not saving facts. It is proving recall, freshness, deletion, and rollback under real development pressure.