Lost in the Middle: Why Your LLM Ignores What You Give It
LLMs don't read long contexts well, not because of capacity limits, but by design. Two studies measure the gap. What it means for how you architect your systems.
10 articles
LLMs don't read long contexts well, not because of capacity limits, but by design. Two studies measure the gap. What it means for how you architect your systems.
The Linux Foundation launches AAIF with MCP, goose, and AGENTS.md. Open governance to prevent ecosystem fragmentation.
A JavaScript API that turns your site into an MCP server. Agents talk to your app through structured tools instead of scraping the DOM.
AI agents have a fundamentally different attack surface than chatbots. A complete map of the threats — from indirect prompt injection to tool supply chain attacks.
How to give an AI agent all the tools it needs to run, analyze, and see your application — and become truly autonomous in detecting and fixing errors.
A deep dive into the review-manager powering opencode-team-lead — how it selects reviewers, isolates their contexts, arbitrates disagreements, and produces a structured verdict without reading a single line of code.
On long tasks, an all-in-one agent loses track and reviews its own code with the same blind spots that produced it. I built opencode-team-lead to fix both problems with a strict delegation pattern.
SHA256 as a KDF, no salt, no versioning — the classic encryption pitfalls in Go, and how cryptio fixes them with Argon2id.
In a codebase that serves as the source of truth, everything an agent needs to 'know' should be mechanically verifiable. Giving your agents memory is patching in prose what should be encoded in the harness.
Anthropic and OpenAI published their harness engineering post-mortems weeks apart. Different problems, same conclusions — here's what to take away.