Post

truly continual (mostly), but if they were? Workflows would be vastly simplified. Shopify: AI judges on specific tasks that their models do, if they don’t outperform humans or frontier models, retrain on new data.

The PyTorch Foundation’s goal is to build models on any chip, any cloud, with any agent. Now Chi Wang formerly of Google DeepMind and now at AG2 to speak about how we can get personal agents to work together.

The “new network” consists of members (people and agents) sharing rooms with their own contexts, with agents able to also work directly together. But agents need accountability and rules to safeguard data.

A theme I’ve seen in a few talks about agent-agent communication is that agents can carry requests from its owners but never their authority. Wang proposes a broker in between agents to do this.

This shows an example workflow for task-focused agents coordinating to report a bug, implement a fix, and approve the changes.

“Coordination could change the scale at which people organize work”

Now we are hearing from Paul Conyngham, the guy who worked with multiple scientists to develop personalized mRNA cancer vaccines for his dog, and is scaling that to develop computationally designed mRNA cancer vaccines.

Already, he is speaking of building ontologies and doing things we’ve done in healthtech and biotech for doing this work - a good thing not covered in any of the popular coverage of what he and his team did. An early idea was “finding the diff” between baseline DNA and cancer cells. Not a new idea,

But he outlines the technologies that converged to allow this to happen: individual sequencing, computational genomics and AI, and mRNA vaccines. He explains that he used AI as a learning partner to learn how these technologies worked and the science behind them.

A lot of this work is, unlike the popular coverage, about what you’d expect for personalized vaccine development. LLMs seem to have largely been used as personal learning assistants. A very good intro to how modern drug discovery works.

Hard to argue with results. His new venture Gamgee, wants to scale this up eventually to humans.

The real thrust of this talk is how to work effectively across disciplines. Emphasizes that the human is the one that has to “be the agent” to make it happen.

Now @ashtom.bsky.social is up to talk about the future of version control for agents, something I know a lot of people here are interested in and working on. He opens to show us how number of PRs is skyrocketing over time and the size of PRs is 3x what it was 18 months ago, causing more changes

accepted without human review. This is bad! But…the SDLC as we know it is dead. It was so focused on making loops intelligible to humans while driving down the size of those loops. But agents are now driving these. So what do we do to make sure our intent travels on the loops with agents?

“Agent session logs are as valuable as code” Although I contend that tools like Deciduous are better. Thomas believes session logs should be committed as git objects. I’m not sure most devs want their full logs exposed, though.

He does point out that we need a wider traceable unit, but I missed that slide. Still, our infrastructure must be decentralized. We are using a centralized tool for decentralized git, we’ve largely lost the advantages of decentralization.

We need infra that is decentralized, independent, and open source. Happy to see that this field still has big believers in the real ethos of OSS - people creating together with no care for national governments, org charts, real names, or affiliation. Inspiring talk.

Next up is Dr. Shazam Laddad from AWS on how to verify code correctness. He introduces “universal properties” - specifications that apply to *any* application regardless of implementation. Strongly typed languages have this, eg rust preventing multiple borrows.

UPs can be defined once and reused, living outside the agent and forming guardrails around agents. Introducing infinity.hydro.run an agent runtime for concurrent agents.

It’s important to find your own universal properties:

Governance (identity, registry, policy, and observability) for all agents inside of an enterprise.

This control plane can allow users to swap models, harnesses, and agents across an enterprise while retaining configs, history, etc. He closes with some best practices for using agents in the enterprise.

Now Kelly Fang on how @anthropic.com is using MCP inside the org. They have built a Claude directory to show all of the connectors and plugins that work across all Claude tools. Unclear if this is related the MCP server directory or a sister tool.