Show Notes
Dan and John Berryman discuss why today’s terminal- and IDE-centered agent harnesses are too narrow. John argues that agents should be able to see and act across the applications, websites, files, and physical environments that make up a person’s life. His Rook project is an attempt to make those contexts addressable while allowing people to keep using the agent harnesses they already trust.
The conversation turns to a future in which websites expose agent-facing capabilities and applications reshape themselves around conversations. John imagines an assistant that can combine a person’s notes and shopping list with location, store inventory, and aisle information. He says feasibility is the major barrier that recently moved into reach, while security, transparency, trust, and standardization remain unresolved.
That tension becomes concrete in the discussion of permissions and sandboxing. John expects people to begin with constrained, read-only access and gradually grant more authority, with dry runs, approvals, and reversibility helping determine when an agent can act on its own. He also points to emerging conventions such as skills files, AGENTS.md, and llms.txt as ways for agents to discover what they can do in a given domain.
John’s strongest practical advice is to replace bespoke workflow code with skills written in plain English whenever the model is capable enough to follow the instructions. He describes building a Zoom assistant and a candidate-vetting workflow this way, arguing that subject-matter experts may eventually be able to read and rewrite the software directly. In his view, the agent is increasingly just a loop around a model and tools, with frameworks such as LangGraph becoming less necessary for many applications.
The discussion of memory is more skeptical. John argues that chunking old conversations and retrieving them by textual similarity does not reproduce the way human experience turns mistakes into procedural knowledge and taste. For now, he prefers explicit, visible review of a completed task followed by packaging the generalized process as a skill. That approach is also his answer to the limits of AI-generated writing: agents can remove ums, edit dead space, and produce polished drafts, but they still do not know what a particular person considers good.
Looking ahead, John expects more agentic applications, just-in-time custom software, and world models that simulate environments for robotics. He is optimistic but clear-eyed about the risks, closing with a plea to direct the technology toward useful and humane outcomes rather than dystopia.
Full episode notes
Click here to view the episode transcript.
Chapters
(00:00) - Why agents should be unharnessed
(03:32) - Rook and context-aware agents
(13:23) - Agent-facing applications and trust
(18:23) - Sandboxing, dry runs, and reversibility
(21:29) - The new programming language is English
(30:28) - Why bespoke agent frameworks are fading
(34:13) - What agent memory gets wrong
(42:50) - Explicit memory through skills
(51:07) - Taste as the final frontier
(57:49) - World models and the future
⠀
Links from the show
--------------------
Arcturus Labs Blog
The AI Product Era You're Building For Might Already Be Over
Unharnessed Agents Power the Future of AI Products
Relevant Search
Prompt Engineering for LLMs
Agent Client Protocol
OpenClaw
Hermes
Pi
LangGraph
Pydantic AI
Tailscale
OpenStreetMap
Model Context Protocol
⠀
Guests
-------
John Berryman, Founder, Arcturus Labs
Website
X
YouTube
⠀
Follow the podcast
-------------------
Threads
TikTok
⠀
Follow Dan Gerlanc
-------------------
X
Threads
Bluesky
More from Agents and Engineers | Agentic AI, Software & Agentic Engineering
Allen Hutchison is the co-founder and CTO of Vycari, building AI agents for everyday tasks with Chris Perry. He spent 21 years at Google, most recently as a Distinguished Engineer at Google DeepMind, helped found Gemini CLI, and built the Antigravity SDK. His earlier leadership work included Google's mapping products, Core ML, and the AI Developer organization. He writes Letters from Silicon Valley. Dan and Allen discuss why consumer agents need familiar interfaces. Vycari embeds agents in mini apps, including a calendar booking tool that works with meeting preferences and buffer time. Allen wants users to think about the assistant's usefulness rather than the technology behind it. Vycari's two founders have landed well over a thousand PRs in roughly a month. Planning documents guide implementation, nightly jobs keep those documents aligned with code, and review follow-ups become issues that agents can implement. Independent reviews and reflection on user trajectories help identify problems, while Allen retains the final merge decision. Their engineering choices favor quick feedback and manageable costs. Vycari is separating microservices into their own repositories and runs Python and Postgres on VPSs. Allen explains how inexpensive models, stronger models writing their procedures, and deterministic integrations support the goal of a $5 monthly subscription. Full episode notes Transcript Chapters
Ben Morss is a Developer Advocate at Mistral. He previously spent 9.5 years at Google as a Product Manager for advanced APIs in Chrome and a Developer Advocate for the web. He earned a bachelor's degree in Computer Science and Music from Harvard University and a doctorate in Music Composition from the University of California at Davis. He continues to make music and write about music theory.
Cory O'Daniel is CEO and co-founder of Massdriver, an internal developer platform and platform orchestrator for governed self-service infrastructure using Terraform, OpenTofu, and Helm. He has spent more than 20 years building teams and startups and working with cloud infrastructure. He created Bonny, an Elixir-based Kubernetes operator framework, and is a co-founder of OpenTofu. Dan and Cory discuss what infrastructure agents need to act safely. Massdriver models environments and resource relationships so agents can query context directly, while separate agent identities and attribute-based permissions constrain their access. Cory describes a staging workflow that tests infrastructure changes through real deployments before he promotes a release to production. Cory explains how a tested, consistent codebase supports his team's Claude Code workflow, why enforceable tooling matters more than long instruction files, and why human review still matters. He argues that engineers should build quality controls and deployment paths for domain experts who can now create software themselves, with collaboration and operational ownership still needing attention. Full episode notes Transcript Chapters
Stephanie Jarmak is an AI Engineer at Omni working on multi-agent orchestration and code intelligence. She is a research affiliate for NASA's Science Explorer (SciX) and a maintainer of Gas City, an open-source orchestration builder for multi-agent coding workflows.
Eric Ma is a Senior Principal Data Scientist at Moderna, where he leads the Data Science and Artificial Intelligence (Research) team. Previously, he conducted biomedical data science research at the Novartis Institutes for Biomedical Research and earned his doctorate in Biological Engineering from MIT. Eric is also an open-source software developer known for leading pyjanitor and nxviz and contributing to NetworkX and PyMC.