MLOps Community

Coding Agents Conference. March 3rd at the Computer History Museum. luma.com/codingagents

Hear the whole thing: go.mlops.community/cjlrjr Spotify: open.spotify.com/episode/24XVJn…

Cost per token dropped ~80%, then reversed. Josh Collier, FinOps Lead at @Superhuman, on why production LLM bills are trending up - model prices doubling and a ~10% US data residency premium teams miss.

Watch here: home.mlops.community/home/videos/wh… Listen on Spotify: open.spotify.com/episode/62rPUX…

MCP is going stateless in its biggest release yet. David Soria Parra, a core maintainer, on why session-based state broke at hyperscaler scale and how stateful bits like elicitation now travel over the wire each round.

go.mlops.community/hcr1qg Where do people land on the workflow first claim? Has anyone seen scattered use cases pay off, or does it fall apart the way he describes?

3⃣ On jobs, Manish says the bottom 25% in a role are the ones AI pressures, while the top half just get faster. Uncomfortable framing, and he stands behind it.

New episode with Manish Dasaur from @PwC, who's spent 20+ years helping companies figure out where their AI money goes and whether it comes back. A lot of it lands on one stat from their research: 88% of companies say they're doing AI, only 33% say they're scaling it.

Watch here: youtube.com/watch?v=-tN2fy… or listen here: open.spotify.com/episode/1G0WQf…

Several models, one ten-sentence story prompt, no context - and they keep returning a lighthouse keeper named Elias. New Debug Mode on why models converge on the same default output, and what it means for harness design and hot-swapping.

ACP v2 is available as a draft. 📝 Highlights: - Updates beyond user-initiated turns - Stream & update messages, tool calls, and terminal output - Structured file changes and better diffs - Improved extensibility and forward compatibility Help shape v2: agentclientprotocol.com/announcements/…

Which camp are you in? 👇 Full session: home.mlops.community/public/events/…

"I don't write prompts anymore. I write loops." "Prompt engineering is dead" is clickbait — but stop hand-verifying every turn and engineer the loop that verifies for you? That's real, and already in prod. Best line from this week's AAIF reading group.

Hear the whole thing: home.mlops.community/home/videos/th… open.spotify.com/episode/6z8Orn…

@jlowin, creator of FastMCP, on Prefab: return a UI component from a tool, and it renders as an interactive React app in the client, with the result going to the user, not the LLM's context window. Token-efficient MCP.

Listen: go.mlops.community/t8xmjw Where do you land on "fast beats right" — fair for most agent decisions, or the thing that burns you first?

3⃣ "Fast beats right" in a lot of moments today, backed by the two-bot refund story, where a satisfaction reward pushes the other bot into handing out money.

@debergalis (CTO/Co-founder of @apollographql ) sat down at the MCP Dev Summit to talk about what happens when you point agents at the pile of APIs a company has stacked up over 20 years.