The AI agent I built just booked me a Zoox through an API I didn’t know existed. "Call me a Zoox and text me updates". I was half joking, but Solid wasn’t. Solid reverse-engineered the app, retrieved my user token, and called the underlying API directly to book me a car. Then it monitored the status and texted me updates, just as I had asked. New AI asks for something more uncomfortable: describe the outcome, leave the method open, and accept that sometimes the human is the constraint. AI is still underrated.

A beta user texted me this morning, raving about our calling agent feature. I thought she was drunk. We haven’t built a calling feature. So I pulled up the log. She was not drunk. She had asked Solid to call local suppliers and check if they have a part in stock. That was the whole request. She did not tell it how, she just gave it the goal. Solid decided the job needed phone calls. Step by step, with no human touching it, here is what it did: • Created its own Google account • Used that account to sign up for Bland, an AI phone-calling service • Paid for it out of its Solid balance by creating a virtual card • Got blocked when Bland flagged its connection as a datacenter bot, so it spun up its own relay machine to operate on a non-datacenter IP address • Picked a British voice, called the suppliers, sat through the phone menus, and spoke with human agents • Reported back with the best supplier There was never a calling “feature”. The agent needed to make phone calls, so it proactively figured out how, the same way a human would. Agents can do so much more than just micromanage us. We just have to allow them to.

We deployed it live at overeasy.dev!

Just gave a talk at @AITinkerers about LLM planning and context, which are both core to Codapt's production-ready code generation. After 16 iterations of our pipeline, we found that planning alone doesn't boost LLM performance. But these techniques do work: 1. Deduplicate redundant context: There’s no need to pollute your context window with multiple reads and writes of the same file. 99% of the time, the LLM only cares about the latest version. Clean up those duplicates! 2. Track context dependencies: Most of the time, agents go through read-write patterns. The writes usually depend on the reads immediately before them, but often the previous reads continue to take up context window space. Clean up the previous reads when you make a mutation. 3. Rollback mechanisms: If your agent is working in a system that can support destructive mutations, utilize that. Take snapshots, and let the agent use them to escape from dead-end paths. Letting the agent mutate earlier and make mistakes (and learn from them) often yields better results than over-populating the context window with too many reads. Thanks @dexhorthy for organizing such an amazing event — always inspiring to connect with fellow builders in this space.