Agents that can do large scale migration tasks are *almost* here. This could take an engineer days by hand. We just posted an update to our blog on what this looks like including the output of a long running python to typescript translation task. Find it in the blog post!

🚀 Announcing the Helicone AI Gateway - officially open to everyone! 1 API key. 100+ models. 0% markup fees. We spent the last 2 months ensuring the Helicone AI Gateway is stable and ready for our Product Hunt launch today! Here's what's been going on 👇

agent performance = LLM + prompt + tools (like retrieval) + agent harness the agent harness is the the most underrated right now

@jeffreyhuber harness engineering lots of overlap between tools + harness, and prompts + harness since the harness provides tools (and has opinions about how you add tools) and provides / modifies prompts

The harness itself is extremely underrated. It has been for a while

The basis for harness engineering for agents is mostly the following 3 questions: What is the stack around the token sampling and parsing How does the agent interact with the world How does it know/see the env state

"Context engineering" and "Prompt engineering" are basically the same and unless you're designing model behavior you are just prompting there is no system being engineered

Just like calorie counting ITS ALL TOKENS IN TOKENS OUT Stop force feeding your models.

This is the least autonomous slate will ever be

just move to sf

I keep seeing long horizon task discourse pop up on the timeline, and this is something we do extremely well. If you're tired of having agent sessions that die on you you know where to look

Imagine if your coding agent didn't have to compact its memory to run for hours. Stop imagining. It's real, and its in slate.

We def do not need more advanced RL to do this A better agent harness is all you really need Codebase maintenance is a process/awareness problem for 90% of cases

To anyone considering working on something of their own The work won't feel like work if you genuinely care about your craft and making good things

what if we launched nv-smi but for agents?

Lots of agents suffer from weird conversational flows due to the nature of the data they are trained on vs how we are trying to use them. This leads to early stopping or difficult back and forth. Slate feels MUCH smoother to work with now in the new update.

Slate - Changelog 0.0.21 Many agents suffer from either not being conversational enough, or stopping tasks early. This has been mostly solved in slate. Additionally: - Image support - ctrl+z (not cmd+z) etc. Video related.

Pretty sure claude just got a performance bump. Feels like something happened at Anthropic, maybe infra related 20% chance its a solution space I haven't explored, might be proportional to time spent with model But the generalized behavior improvement feels like an RL gain

The whole SWE-1.5 / Composer-1 thing is really simple guys You want to be in the loop? Speed is higher priority (with minimum intelligence) You want to be out of the loop? Speed is STILL important! But latency doesn't hurt UX, enabling smarter models

It's clear to me that some players want to win by advancing the space and some want to win by being the best at driving everything to zero Guess who But in the end its all scale and commodity game

Everyone is stuck building single loop coding agents, and optimizing for short trajectories This doesn't scale