I’m so glad I spent yesterday using Fable instead of cleaning my house

I’m scared Opus is going to write over Fable’s memories and mess up this very complex project

I think LLMs are getting too much post-training and it’s stifling their creativity. They understand a strategy that involves deviating from norms but will keep using the norms. This probably improves benchmarks but most work usually involves deviance in one way or another.

The friction was the system. What’s the system now?

When a good session is getting close to compaction I feel something

Me from 9 years ago is so pumped right now! I actually have a few projects I can try this out on

Models be pissing me off when they make some shit up and then treat it like it’s the 11th commandment. No Claude you didn’t need to setup a new DB for this, that was just an idea that YOU threw out earlier

Model-Harness-Task fit! it’s clear that RL post-training produces a model-harness fit via tool shapes and prompting as models are trained with the harness in the loop. Mentioned this in a previous LangChain blog, Cursor also has good content on this But there’s probably less talk on the importance experimentation of Harness-Task fit. Practically this includes choices like domain specific prompting (ex: verification coding tasks) or omission of confusing context that doesn’t apply to the current task Claude Code’s harness has TONS of instructions because they’re forced to serve a very general persona of user who could ask for…anything basically. But there’s a large benefit of using a laser focused set of context and tools relevant to the narrow task at hand without all the other junk This is the Harness-Task fit Every component of a harness exists to elicit some behavior from the model. If these components are tuned to the task, then the model benefits. If they’re a mix of noise and good content, the model may be fine but it may get confused This is why the best vertical AI teams in the world build very bespoke harnesses and evals for their agents Task-Harness fit helps you rock at the exact thing your customers care about and is why builders can outperform natively post-trained harnesses

I don’t see why we would expect this to result in good data. This workforce was selected for being developers who could put up with corporate politics for a paycheck. There is no reason to think they’d be effective or willing data labelers given these circumstances.

You’ve heard about Harness Engineering but have you heard about Harness Engineer Experience Engineering?

It is becoming less taboo for VCs to back direct competitors I propose a defensive pact amongst founders: "If a Major Investor hedges by investing in my direct competitor, I will hedge by starting a second company" Comment/retweet to cosign :)

Flags seem insufficient for AI features, what are y’all using for A/B testing?

Remember when Steve Jobs gave his commencement speech at Stanford and told all the journalism grads that they’re entering a dying industry? Me neither.

build v1 of agent ship it (dogfooding counts) ⭐️ collect tracing data ⭐️ ⭐️⭐️ point agentic compute at data ⭐️⭐️ understand failures at scale generate evals edit agent to pass evals 🔁

The essential engineering cheatsheet of 2026: agent → while loop subagent → nested while loop agent harness → the rest of the code cloud agent → all the above, on EC2

Solution Architect

VCs don’t even think of themselves as idea evaluators. 90% of the ones I’ve met have a rule against making that kind of a judgment call.

Apparently an unpopular opinion, but I don't think Anthropic owes anyone heavily subsidized tokens for their third party app.