Artwork for Agents and Engineers | Agentic AI, Software & Agentic Engineering

Agents and Engineers | Agentic AI, Software & Agentic Engineering · August 18 · 1 hr 5 min

The Handoff Packet for Local Agentic AI

Paul Everitt is a Python and Web Developer Advocate at JetBrains. He previously co-founded Zope Corporation, helped bootstrap the Python Software Foundation and Plone Foundation, and served as a US Navy officer who started www.navy.mil in 1993. Paul and Dan discuss how local agentic AI could reshape software engineering, from smaller models supported by static analysis and better tool calling to workflows built around specs, tests, and focused handoff packets. They compare local, cloud, and hybrid model setups and examine what happens when agents work with less supervision. The episode's titular handoff packet sits at the center of a three-level workflow. A large model understands the project, a medium model creates an execution plan, and a small local model carries out a tightly scoped task. The packet gives that smaller model the context, constraints, tools, and success criteria it needs. They also discuss spec drift, Markdown debt, community-maintained Python training data, and the responsibility developers retain to read, test, and own the code they ship. Full episode notes Transcript Chapters (00:00) - A profession changing by the hour (02:54) - Why Python still matters with humans in the loop (04:37) - Static analysis for small local models (11:51) - DwarfStar and the local inference stack (17:28) - Apple's on-device AI push (23:29) - Why unsupervised agent runs fall apart (28:46) - The big brain, medium brain, little brain (41:19) - Spec drift and Markdown debt (47:45) - Pyrefly and duck typing for agents (51:22) - Better tool calling for local coding models (01:00:10) - Community-maintained Python training data (01:03:42) - Bring back the engineering ⠀ Links from the show -------------------- arXiv Pydantic AI LM Studio OpenRouter DeepLearning.AI spec-driven development course Spec Kit Superpowers Pyrefly Mellum Flower Liquid AI ⠀ Guests ------- Paul Everitt, Python and Web Developer Advocate at JetBrains Website LinkedIn DeepLearning.AI: Spec-Driven Development with Coding Agents ⠀ Follow the podcast ------------------- LinkedIn Threads Instagram TikTok ⠀ Follow Dan Gerlanc ------------------- X LinkedIn Threads Bluesky

0:00 · A profession changing by the hour-1:05:38

Play in

Transcript file

transcript

No transcript — this publisher did not publish one.

show notes

Paul and Dan discuss how local agentic AI could reshape software engineering, from smaller models supported by static analysis and better tool calling to workflows built around specs, tests, and focused handoff packets. They compare local, cloud, and hybrid model setups and examine what happens when agents work with less supervision.

The episode's titular handoff packet sits at the center of a three-level workflow. A large model understands the project, a medium model creates an execution plan, and a small local model carries out a tightly scoped task. The packet gives that smaller model the context, constraints, tools, and success criteria it needs.

They also discuss spec drift, Markdown debt, community-maintained Python training data, and the responsibility developers retain to read, test, and own the code they ship.

Full episode notes

Transcript

Chapters