Launching today
Case by DaemonLabs
Stop dreading the desktop part of your AI agent.
15 followers
Stop dreading the desktop part of your AI agent.
15 followers
Most AI agents drive desktop apps by clicking around and hoping. Case replaces that with typed, verified procedures: requires, ensures, structured failures, idempotency labels, and reliability stats pinned to a specific app version. 50+ procedures live for DaVinci Resolve today, with Photoshop and Logic shipping next. Drop into Claude Code, Anthropic Computer Use, or Hermes. The desktop part of your agent stops being what breaks.
Hey PH, Ishant from Daemon Labs.
I chose GPU compute over rent. At 20. In college. And I'd do it again.
While my batchmates were grinding interview prep, I was reading transformer papers at 2am in my hostel room.
No lab. No funding. Just a bet that AI was the most important thing happening in the world and I needed to be inside it, not watching.
We applied to Founders Inc's Canopy program. Got rejected. I didn't take it. DM'd partners on X. Showed up in their threads with real responses. Sent handwritten letters promising a specific deliverable by demo day. Hours before the program started, we were in.
Same year, my paper on procedural memory in language agents got into the ICLR 2026 MemAgents workshop. Building that benchmark made one thing obvious:
Agents aren't failing because they can't reason. They're failing because the execution layer is a complete disaster.
Silent crashes. Pixel-dependent clicks that break every software update. Zero structure in the failure signal. Every retry is a gamble. Every demo is a prayer.
That insight became Case.
A verified procedure runtime for the desktop apps your agent has to drive. Typed calls. Structured failures with retry flags and recovery docs instead of cryptic stack traces. Idempotency labels so retries don't compound bugs.
Today we're shipping two things:
case-sdk (open source): the procedure runtime. Build your own procedures to operate any desktop software. → https://github.com/DaemonLabsInc/case-sdk
case-api (managed): plug-and-play access to our verified, battle-tested procedures. DaVinci Resolve today, more apps shipping soon.
99.6% reliability measured across 51,000 real customer calls. A number you can actually ship to your own customers.
We started with DaVinci Resolve because that's where the pain is loudest. The whole desktop is next.
Everyone from PH gets 50 procedure runs per day, free, for 7 days.
If you're building agents that touch real software, what's the procedure that's killed your demo? 👇
Couldn’t agree more with this @ishu86,
This is basically what got me interested in Case from the start.
I’d already seen how fast things fall apart once agents leave demos and have to deal with real software. The models were improving fast, but the execution layer still felt fundamentally unreliable.
Random silent failures. Retries digging the hole deeper. One tiny UI change nuking an entire workflow overnight.
Most people were just adding layers around the chaos instead of fixing the core reliability problem. We treated this like infrastructure from day one, not another thin AI wrapper pretending the hard part doesn’t exist.
Glad we are finally working on this with people who see the problem the same way.
Hey PH, Pulkit from Daemon Labs.
Built Case because we kept hitting the same wall with AI agents. The thinking part was great. The doing part was a mess.
A button shifts. A click doesn't register. The agent retries forever. And the workflow that ran clean yesterday just... doesn't today. You end up babysitting the thing instead of trusting it.
I didn't take the normal route into this. Spent most of my time deep in systems, infra, and applied AI, the unsexy stuff. And the more I poked at agents, the more obvious it got. The models are smart enough. They've been smart enough for a while. What's missing is everything underneath them actually working when it has to.
So we started building Case.
Instead of agents fumbling around the screen guessing where to click, Case runs procedures that are structured, typed, and check themselves as they go. They know what app version they're on. They don't break when something moves. They actually finish.
Early days. But it's the thing we kept wishing someone would build, so we're building it. 🚀
@pulkit_aggarwal5 Sounds about right. It’s honestly crazy how many people expect deterministic behavior from systems that are fundamentally probabilistic.
Most of the real work ends up being about the layers around the model, the harness, the recovery logic, the guardrails, all the stuff that actually makes execution stable.
Getting that right was hard, which is exactly why we ended up open sourcing a lot of the infrastructure behind it.
Hey PH, Mayank from Daemon Labs.
With all the hype around computer-use agents and software automation, I started exploring the space myself. And almost everything I saw boiled down to agents taking screenshots and guessing what to click next.
That felt fundamentally wrong.
If a system is meant to automate software reliably, it shouldn’t just look at the screen. It should understand the software, know the available actions, and handle every workflow inside the app.
That led us to a simple idea:
software automation should run on procedures, not guesses.
Instead of treating every interaction like a vision problem, we started building structured procedures for software itself. Reliable flows that know how to operate an application, recover from failures, and execute deterministically.
That eventually became Case.
For the PH crowd: 50 free procedure runs on us. Drop your email at daemonlabs.run/ph, key lands within 24h, then point case-sdk at DaVinci (or any other macOS app) and drive it from Python or Claude Code.