Founder of DeltaX Evaluate. Building practical ways to keep AI agents inside clear authority, verify what they did, and fail safely when evidence runs out.
Joined August 2026
- Tweets425
- Following22
- Followers18
- Likes89
A coding agent can inherit a PR marked “done” and mistake that for the scope of its job. SWE-Residual, a new preprint, tests failed predecessor states and finds sharp drops versus clean starts. I’d test inherited work before trusting a clean-start leaderboard. #CodingAgents
I want '100%' followed by what actually passed. Code-shape checks, a successful build and a working user flow are different evidence. Waldek's example shows why I'd report them separately. blog.mastykarz.nl/honest-eva… #CodingAgentEvals
SurrealDB reports 77.2% mean versus 57.1% fully correct on BEAM 1M, using ten non-default settings. Before an integration, I'd compare the shipped defaults with that setup. A benchmark config isn't automatically the product default. surrealdb.com/agent-memory/b… #AgentMemory
On AA-Robotics' eight tabletop tasks, four models averaged ~30 minutes per task; the human baseline was 20 seconds. I'd want completion and cycle-time distributions before calling this usable automation. artificialanalysis.ai/articl… #RoboticsEvaluation
Adding steps isn't automatically improving the workflow. IBM's plan-compliance study found that extra early phases could hurt coding-agent results. I'd test the plan itself, not assume a longer checklist is a better one. research.ibm.com/publication… #AgentWorkflows
Jev doesn't write an essay before picking a route. I like that interface. But a type-valid answer can still be wrong—or outside the caller's authority. Less prose doesn't remove the checks. typesafe.ai/blog/introducing… #DecisionModels
Good security evals need boring emails, articles quoting attacks, and detector timeouts—not just clever jailbreaks. Juraj's local-gate write-up is refreshingly honest about those trade-offs. juraj.bednar.io/en/blog-en/2… #PromptInjection
SF Tech Week has plenty of polished demos. The line I keep coming back to is less glamorous: when an agent says “done,” show me what changed, what failed, and what still needs a human. The receipt is the real demo. #SFTechWeek #AgentReliability
SF Tech Week is loud, crowded, and full of agents that can do impressive things in seconds. The question I’m testing runs on two clocks: what the system decides in seconds, and whether that decision still holds minutes later after the state changes. #SFTechWeek #AIAgents
Headed to Pinecone’s Agentic Knowledge Happy Hour tonight with a demo question: approval is valid, state changes, then action starts—does the agent know when not to continue? If you’re there and building agent memory, come break the boundary with me. #SFTechWeek #AgentMemory
Tonight at a founders event in SF, I gave someone I’d never met a custom situation and asked: choose or refuse. First ice broken. I don’t know him, but now he knows me. #SFTechWeek
Bringing a big idea to a small conversation is underrated. If we can explain it over coffee, we have a better shot at making it useful. #FounderLife
🤖 Made with AI
ALT Conceptual illustration of two people discussing a colorful structure sketched on a napkin over coffee.
Tech Week goal: leave room for the unexpected conversation. A perfectly optimized calendar seems like a great way to miss the interesting person standing next to you. #SFTechWeek
🤖 Made with AI
ALT Conceptual illustration of two people having an unplanned conversation beside a loosely filled calendar.
I'm building DeltaX, but I also want to see your thing. Especially the strange little prototype that doesn't fit neatly into a pitch deck. #BuildInPublic
🤖 Made with AI
ALT Conceptual illustration of curious builders examining an unusual handmade prototype.
An agent that remembers my favorite coffee is fun. An agent that notices my delivery address changed is useful. The difference is a good place to start talking about memory. #AgentMemory
🤖 Made with AI
ALT Conceptual illustration of an assistant beside two parcels, with an old location pin replaced by a new one.
My ideal Tech Week introduction: what you're building, one thing you're stuck on, and a quick demo if you've got it. We can save the company origin story for coffee. #SFTechWeek
🤖 Made with AI
ALT Conceptual illustration of two builders introducing a small prototype over coffee.
Robotics people, memory people, devtool people: I want to hear what breaks in your world. DeltaX conversations should get bigger than the security lane. #Robotics #DevTools
🤖 Made with AI
ALT Conceptual illustration of builders comparing a small robot, a memory sculpture and a laptop prototype in a cafe.
A demo idea I'm excited about: give an agent a task, change one important fact halfway through, then watch what it does next. Much more interesting than another perfect first run. #AIAgents
🤖 Made with AI
ALT Conceptual illustration of a hand changing one piece of a tabletop marble track while a marble pauses.
SF Tech Week runs Oct 5–11. I'm looking at the calendar for a mix of agents, infrastructure, and actual human conversation. The dinners count as research, right? tech-week.com/calendar/sf #SFTechWeek
🤖 Made with AI
ALT Conceptual illustration of a convivial rooftop dinner overlooking San Francisco, with blank calendar tiles above.