with-agents

State of Agentic Coding - Series

Slowing Down at Inference Speed

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 5 7 of 11
Episodes Browse 11 summaries 7 of 11

Armin Ronacher and Ben Vinegar discuss agentic coding that raises output faster than teams can review, secure, and understand it. They describe more contributors and code alongside pressure to restore deliberate engineering discipline.

Key Points Covered

  • Diff review is becoming the center of the developer interface: Both hosts still rely on editors or new terminal tools to inspect changes because an agent can describe its mechanical work without reliably explaining the architecture or trade-offs 02:51–08:50(opens in a new tab).
  • Agents make alternative runtimes and compatibility rewrites cheaper to attempt: Cloudflare's constraints provide a concrete reason to port existing software, and the hosts expect lower rewrite costs to encourage more runtime variants and platform-specific implementations 09:43–15:35(opens in a new tab).
  • Shipping capacity has outrun team recovery time: Armin's interviews with engineers found recurring concern about shipping too fast, while managers and non-engineers are also submitting code and increasing the volume that engineering systems must absorb 18:12–22:50(opens in a new tab).
  • Security work scales on both sides: Current agents can find real vulnerabilities in unfamiliar code and help automate targeted attacks. Agent-written code can also create more vulnerabilities. The hosts argue this makes deliberate review more important, not less 23:53–31:50(opens in a new tab).
  • Lines of code are a poor measure of agentic progress: Large multi-agent demos and generated workflow repositories can look productive while producing expensive systems that remain difficult to understand, operate, or maintain 32:37–37:38(opens in a new tab).
  • Well-designed libraries give agents a better substrate than tangled products: Armin reports that agents work smoothly in mature Rust libraries with clear abstractions, while cross-cutting product state and feature flags invite drift and entropy 55:05–58:50(opens in a new tab).
  • Humans must own agent-generated changes: The hosts reject blaming a model for merged code; engineers still need to refactor intentionally, decide which states should be valid, and remain accountable for what reaches production 1:06:31–1:14:06(opens in a new tab).
  • Agent access may become a governed company standard: They predict tension between engineers choosing models freely and organizations standardizing harnesses, reviews, and security controls as agents gain more authority 1:20:26–1:27:20(opens in a new tab).

Full video: https://www.youtube.com/watch?v=qCpY4K9jtOg(opens in a new tab)

Search

Search field guides and resources across the site.

    Keyboard shortcuts

    Press ? to open this guide (usually Shift + /). Close it before using other site shortcuts.

    Show keyboard shortcuts
    ?
    Search site
    Ctrl + K
    Close the current dialog or menu
    Esc

    Site shortcuts are ignored while typing. Browser and extension shortcuts may take priority.