Post

Simon Willison
@simonwillison.net

If you're interested in the nascent field of context engineering (the new, more cromulent alternative to prompt engineering) Drew's piece here provides some excellent nomenclature for some of the challenges you might face...

Drew Breunig@dbreunig.bsky.social

More context, more problems. www.dbreunig.com/2025/06/22/h...

dbreunig.comHow Long Contexts FailTaking care of your context is the key to building successful agents. Just because there’s a 1 million token context window doesn’t mean you should fill it.

11 reposts104 likes

John Wards
@johnwards.bsky.social

I’ve been noticing this context confusion while using agent mode in copilot this last week. After a pretty long session it just starts getting confused. I’m starting doing summary docs when I get to a good phase point, then using that to start a new session. Quality is much higher.

John Wards
@johnwards.bsky.social

Oh talking of quality. Good god open ai 4.1 is so shit in copilot it’s like they want to make me use the premium models. Basically refuses to do work, does it half arsed. Is it just me??

Simon Willison
@simonwillison.net

... and his follow-up describes a bunch of useful patterns for taking on those problems bsky.app/profile/dbre...

Drew Breunig@dbreunig.bsky.social

And here's how to avoid or fix your context issues. www.dbreunig.com/2025/06/26/h...

dbreunig.comHow to Fix Your Context6 tactics for fixing your context and shipping better agents. As Karpathy says, building LLM-powered apps means learning to ‘pack the context windows just right’—smartly deploying tools, managing info...
Michal Migurski
@mike.teczno.com

These are both great. I’ve been thinking for a while that they map onto 90s/00s-era IA observations about cognitive science that became normal best practices for managing user attention in web interfaces. Breadcrumb nav, hierarchical menus: all wayfinding aides for limited-context machines like us.