What can a bunch of students build when you give them access to Chalk and 36 hours? This past weekend at @Hack_Harvard, we challenged students to create a data-driven agent built with Chalk, and prove that it works. The results blew us away. Here’s a look at how teams used parts of Chalk: Sandboxes: Trace, our winner, ran its wildfire and storm-surge alert loop 24/7 in a Chalk sandbox. Sandboxes hosted its web app and its text-and-call service. Model Gateway: FEEL, our runner-up, sent camera frames to a vision model through a locked-down key: two models, prompt logging off. Scaling groups: Rubber Ducky’s AI math tutor demo ran live on a Chalk scaling group. Resolvers: Antibody hosted its security orchestrator on one and scored live requests with Chalk resolvers. Online and offline queries: FEEL used a Chalk online query that plans a walking path from the current scan plus past trips through the same building and turned each visit into a room-map feature kept for 60 days. Thank you again to HackHarvard and to every team that built on Chalk. A full recap of the event and our winners, with architecture details for each build, is below.

Agents now access sensitive APIs, ship code, and run for days without a human checking each step. Knowing they're secure, doing the right thing, and not overwhelming your systems is the hard part. In Part 1 of the first Chalk Talk, our co-founder @Derander sits down with @NancyZWang, CTO of @1Password, to talk through running agents in production. They discuss: → Securing agents, starting with identity → Writing evals that show whether an agent actually did the right thing → What agent traffic is doing to infrastructure Thanks to Nancy for the great conversation. Part 1 is below.

Day 1 of @AIconference is live, and the exhibit floor is already buzzing. Come find us at Booth 262, grab our latest swag, and see if you can solve our new math proof competition. Plus, we're giving away a Nintendo Switch 2 to one lucky booth visitor. Tomorrow, Theater 2 is the place to be. At 1pm PT our co-founder @marx_elliot takes the stage with "Two Halves of Inference to Build, Eval & Run Agents". More details below. See you on the floor.

Next week is @AIconference and the Chalk team is ready to show up big. Our co-founder @marx_elliot will be presenting "Two Halves of Inference to Build, Eval & Run Agents" on 10/1 at 1pm PT. If you struggle with stale data and evals that can’t predict production performance, this is the talk for you. Plus, the Chalk team will be hanging out on the exhibit floor throughout the conference. Stop by Booth 262 to chat with our FDEs, catch live demos, and grab some of our latest Chalk swag. Schedule time with us in advance by visiting the link below. 👇

Better agents require better evals. Better evals require better context. Most agent evals are broken before they even start. Not because the model reasoned poorly, but because it was scored on a prompt built from stale, disconnected data. That's not a model problem. That’s a data assembly failure. The fix? A unified context layer that delivers real-time data at inference. Next month our Co-Founder Elliot Marx will host a live webinar to show you how to turn your data into actionable context for your agents, so you can run proper evals and deploy with confidence. October 8th at 10am PT/1pm ET. Register at the link below.

What can a hand of poker teach us about how real-time context changes agent behavior? Our Developer Advocate @rkundy built a poker agent to show how recommendations change with fresh data vs historical data alone. First presented at the @temporalio's 'Durable AI' meetup, Rishi now shares his learnings in our new blog post, "Context Has a Timestamp". Link in the comments.