Episode 005 - Evan Barnard, CPO // AgentGrail

Most AI assistants will do anything you ask, and that is exactly why they cannot do anything on their own. Nir, founder of Catch and a former early-stage investor at a Tel Aviv VC firm, drew a frame around his product and told users plainly what sits inside it: calendar, email, travel, and daily operations. Anything else is somebody else's job. On this episode of Agent Hour he explains why that refusal is the whole design. Because Catch only handles admin, it can build a profile of how long you like your meetings, which days you work from home and who your key relationships are, then act before you ask. He walks through the two-key trust model the product runs on, where the agent takes an action unsupervised only when you have granted it and the system's own confidence is high enough, and he describes what happened the night his agent booked over his plans and he refused to reschedule, so the team would feel the cost of the failure themselves. 🎧 Brought to you by Anchor Browser — cloud-native browser infrastructure for AI agents.

Kush Bhuwalka wants to build an agent that starts and runs its own company. First he'd like people to stop running agents on a Mac Mini under a desk. In this episode of Agent Hour, Gabi Weinberg talks with the founder of Puffle AI, an AI teammate working inside Slack, GitHub, Linear and Notion for construction, manufacturing and logistics companies about why always-on machines are the wrong foundation. Keep a box running and every credential, tool and database connection ends up on it, all reachable by the model runtime. Strip those out one by one, as most teams eventually do, and the only thing left is the runtime itself: serverless, arrived at by accident, the same path the web took from rented servers to functions. He's equally direct about the limits autonomous pull requests are "slop," he says, even though Puffle opens them for his own team every day. The difference is whether a human reads one before it ships. 🎧 Brought to you by Anchor Browser — cloud-native browser infrastructure for AI agents.

More than 60% of the traffic hitting Fly.io's properties, and over 70% of the behavior inside the platform itself, is driven by agents rather than humans, and roughly two-thirds of the revenue is now agent-native. Gabi Weinberg talks with Scott Johnston, the new CEO of Fly.io, who was at Netscape for the birth of the commercial web and ran Docker through the container wave, about what those numbers do to an infrastructure company. Fly is seeing single node, single geo, stateful, spiky traffic that has nothing to do with Monday through Friday, and billing per agent per human rather than per application. Scott walks through what Fly built for that, including checkpoint and restore fast enough to roll back a loop that went wrong, and real-time credential injection so an agent can reach its tools without ever holding the keys to them, plus spin-to-zero pricing, the moat question, and whether any of this is a bubble. 🎧 Brought to you by Anchor Browser — cloud-native browser infrastructure for AI agents.

Most browser agents take a screenshot and guess where to click. rtrvr.ai doesn't take screenshots at all. In this episode of Agent Hour, Gabi Weinberg sits down with Arjun Chintapalli and Bhavani Kalisetty, the ex-Google engineers who bootstrapped rtrvr.ai from a Chrome extension into a cloud browser platform and an in-product onboarding agent. Their argument is technical: language models are language models, extraordinary at semantic understanding, mediocre at finding a pixel coordinate. So rather than train a vision model, rtrvr.ai compresses a page's DOM, which can run into millions of tokens, into a representation that preserves every possible action, with element IDs mapped to real on-page text so the agent has far less room to invent things. It also runs on any off-the-shelf model, including text-only ones like DeepSeek; cheap enough, they argue, that a single ad impression could cover an entire agent task. 🎧 Brought to you by Anchor Browser — cloud-native browser infrastructure for AI agents.

When an agent's session ends, its context usually goes with it — and that problem multiplies fast once you've got multiple agents running at once, each with different access to different data. Vas, founder and CEO of Cognee, joins Agent Hour to explain why simple vector search isn't enough for real agent memory, and how knowledge graphs let agents actually reason over relationships instead of just retrieving similar chunks. He walks through Cognee's ECL pipeline (Extract, Cognify, Load) for turning messy multi-source data into something agents can navigate, plus how granular, per-agent access control keeps each agent's "world model" clean as teams scale.