What if your agent's code had to prove it was correct before you reviewed it?
Nada Amin and Fernanda Graciolli of Midspiral at AI DevCon NYC: write the contract, let the agent iterate, ship what the verifier accepts.
Full talk on YouTube: https://t.co/jDyqVL6HXk
Microsoft's Daniel Meppiel says developers will soon write skills instead of code, and the software factory that runs them will need versioning like any other software. The catch: the teams he works with still aren't rigorously evaluating their skill changes. Simon Maple sits down with Daniel at AI DevCon NYC 2026.
Watch the full episode - link in bio - or listen wherever you get your podcasts
We used to build the car. Now we're building the factory.
Skills define good, loops learn from what they observe, and factories get built around business outcomes. With real data from Tessl's factory.
@guypod's opening keynote at AI DevCon NYC: https://t.co/0mCOYLDkEU
Is "agentic workflow" a contradiction?
@aroussi thinks so. A workflow is a deterministic set of steps with known inputs and outputs. Agentic means non-deterministic. Call a workflow agentic and you're misleading people or wasting tokens.
Full episode: https://t.co/UL9pNbGmZr
Paul Stack of Swamp says ops teams can hand their toil to software automation, because agents love the mundane work operators hate. The catch: the people who know the system still decide what happens next.
Watch the full episode - link in bio - or listen wherever you get your podcasts
The first one sold out. AI DevCon London is back on 7 to 8 June 2027, at The Brewery, Moorgate.
Two full days of honest takes and real stories from teams building with agents. Early-bird tickets: https://t.co/waHIOcSFMx
Your agents ship more code every week, and you're still running to keep up.
Ask what each loop leaves behind. Some keep today's work moving, some make tomorrow easier, and some finally test the idea you keep postponing.
@patrickdebois : https://t.co/XaYHvEB7V5
Paul Stack helped build Terraform, but he has swapped infrastructure as code for an agent harness and says he'll never write another line of it. His team at Swamp has moved its production cloud four times in six months, data included. Guy Podjarny sits down with him at AI DevCon NYC 2026 to find out how.
Watch the full episode - link in bio - or listen wherever you get your podcasts
Agents now author 22% of merged PRs at the median company. Only 0.6% get merged by agents without a human review.
A factory where every unit waits for a human inspector is just a workshop with more conveyor belts.
@arcolano (@Jellyfish): https://t.co/iQpheUWSSN
Your agents keep re-deriving the same API, every session. The fix isn't a sharper prompt. It's the loops around your agents.
@sjmaple and Maria Gorinova on feedback loop engineering: https://t.co/tnG7gzQMbL
Independent tickets don't guarantee independent changes.
Ten agents on one repo means collisions, stale assumptions, and review agents talking in circles. Lessons from scaling Tessl's factory.
James Moss (@tessl_io) at AI DevCon NYC: https://t.co/WmkCtSmsq2
Two weeks ago, 700+ engineers came to Brooklyn for AI DevCon NYC. Missed it? Every talk is now on YouTube.
Next up is London, 7 to 8 June 2027, at The Brewery, Moorgate. Save the date and grab an early-bird ticket: https://t.co/waHIOcTdC5
Your software factory works. Then the first five problems show up: agent IT, code quality, repetition, and test quality among them.
The fix each time is to improve the process, not the code.
@drufball at AI DevCon NYC: https://t.co/evOYyeFNXI
If your team only takes smart bets, it isn't taking enough bets. LaunchDarkly CEO Edith Harbaugh on why she encourages dumb bets, and why AI makes it faster to build a feature than to argue about whether it's worth building.
Watch the full episode at https://t.co/goahxTQdSJ or listen wherever you get your podcasts
Worried agents are doing all the work? @Steve_Yegge: "Don't think about it that way. Think Death Star, right. Or whatever your North Star is."
Full conversation with @drufball: https://t.co/bBJV6ujjdG
In a software factory, the code owner didn't write the code. Maria Gorinova explains Tessl's steward role: someone responsible for a repo who makes sure the dark factory works well there, without typing the code themselves.
Watch the full episode at https://t.co/DHUIkOBwC1 or listen wherever you get your podcasts
If an agent writes the code, who do you let merge it?
Some teams ship after a glance at the diff. Others want two reviewers. Same tools. @mreferre (@AWS) says the gap is trust, and only with trust can you build a factory that keeps agent costs in check.
https://t.co/StwugYN1ZQ
Loop engineering means designing the cycle an agent works inside, not the prompt that starts it. Lada Kesseler gives her agent one step at a time, makes it write the result to a file, then makes it read that file back and judge itself. She calls it a centrifuge. She also trusts AI-written tests less than AI-written code, and thinks most teams chasing a software factory are about to drown in garbage.
Watch the full episode at https://t.co/Cl59wRHy5a or listen wherever you get your podcasts