First screen, captured 17 hours ago. Hover to scroll the page, or open the video.

The State of Context Engineering

What HN thinks

A support group where everyone confirms, one by one, that their AI gets dumber the longer they talk to it.

Long context degrades model quality
Many commenters report models getting dumber as context grows and manually reset, write handoff files, or cap context around 200-400k tokens. A few say newer models hold up until 700k or more.
Organization matters, not just length
Placing related information close together (dependency length minimization), progressive disclosure, sub-agents and checkpointing into artifacts help models reason better.
Compaction and caching trade-offs
Pruning mid-context breaks prefix caches, so people debate provider-native compaction, custom harness plugins, and timing compaction around cache expiry.

Consensus: Most agree that context size and quality strongly affect agent performance and that deliberate forgetting or resetting helps. How best to automate this is unsettled.

Recommended in the comments

pi-optchat ↗ (tool): It keeps recent messages detailed and older history summarized, while letting the agent navigate back to the full text.
Codex (tool): One commenter said it manages its own context well enough that they rarely think about window size.

Summarised by Claude from the comment thread 6 hours ago.

Points: 227
Comments: 57
Best rank: #4
Posted: 2026-10-25 by pchristensen
Category: Ai ml

context-engineering llm ai-agents long-context dspy agntcon

All captures:

Full page

Captured 17 hours ago. Click for the whole screenshot. The page is longer than 25,000px, so the screenshot stops there: read the rest on dbreunig.com ↗

Looks similar

Screenshot of nlp.elvissaravia.com: Context Engineering Guide

Context Engineering Guide (nlp.elvissaravia.com)

56 points by Bogdanp 1 year ago | 18 comments

Screenshot of mbleigh.dev: Context engineering is sleeping on the humble hyperlink

Context engineering is sleeping on the humble hyperlink (mbleigh.dev)

161 points by mbleigh 1 year ago | 62 comments

Screenshot of philschmid.de: The new skill in AI is not prompting, it's context engineering

The new skill in AI is not prompting, it's context engineering (philschmid.de)

832 points by robotswantdata 1 year ago | 467 comments

Blog post by Philipp Schmid arguing that context engineering, supplying LLMs with the right information and tools, matters more than prompting for AI agents.

context-engineering prompt-engineering ai-agents llm