State of Agentic Coding - Series
Harnesses Become the Product
Armin Ronacher and Ben Vinegar track agentic coding through monthly conversations about models, harnesses, software quality, open source, infrastructure, economics, and the changing role of developers. Episode 11 brings in Geoff Huntley to discuss durable agent execution, modifying running software through an LLM, the difficulty of inspecting ad hoc tool code, and experiments made affordable by subscription pricing.
Armin Ronacher Episode 2 10 of 11
Episodes Browse 11 summaries 10 of 11
Armin Ronacher and Ben Vinegar discuss subscription economics, model working styles, and programmable agent harnesses.
Key Points Covered
- Subscription pricing changes how freely developers use agents: The speakers contrast heavily discounted model-provider plans with metered tools. They repeatedly qualify their Cursor comparison as perception and secondhand experience, not a definitive price audit 06:36–12:25(opens in a new tab).
- Model preferences emerge through sustained use: Armin describes Codex as suited to long, largely unattended runs and Opus as more conversational. Both speakers say understanding those differences took weeks, not launch-day impressions 14:22–19:07(opens in a new tab).
- Harness design remains a meaningful source of differentiation: Armin defines the harness as the loop, tools, and prompts around a model. He points to Amp's editable handoff and growing cross-tool support for skills as innovation outside the base model 20:10–25:20(opens in a new tab).
- Metered cost can constrain otherwise useful autonomy: Ben says awareness of Amp's token spend made him avoid longer tasks, even though he valued its handoff interface and curated model experience 26:11–28:00(opens in a new tab).
- A minimal harness can become a programmable canvas: Pi exposes a small toolset and lets Armin add extensions that change its interaction model, including custom question UIs and rendered diagrams 32:09–37:51(opens in a new tab).
- Branching preserves good context without preserving wasted work: Armin rewinds to an earlier message, discards an unproductive path, and resumes with new guidance. He also reuses a shared context point to refine several separate tickets 32:09–34:15(opens in a new tab).
- Agents improve when software exposes its state: For a game project, Armin first had the agent build screenshot and state-dump tooling. When those debugging tools failed, he branched back to repair them before retrying the feature 39:56–43:01(opens in a new tab).
- Design for the model's demonstrated strengths: The speakers recommend asking an agent what information or interface would help it, then testing that answer. They favor representations such as files or familiar code when those produce more reliable behavior 43:01–47:04(opens in a new tab).
Full video: https://www.youtube.com/watch?v=xoynR-hWNZY(opens in a new tab)