Addy Osmani (ex-Google Cloud AI/Gemini director, Chrome DevTools lead) announced his Pragmatic Engineer podcast conversation with Gergely Orosz covering engineering roles, DevTools-to-AI agents, and cognitive debt/surrender @addyosmani. Agent-relevant episode segments: "Working with agents" (1:03:03), "Loop engineering" (1:05:52), "The changing role of the software engineer" (1:12:55), and "How Addy uses AI in writing" (1:18:15) @GergelyOrosz. Osmani's take in the episode: engineering/product roles will "unbundle" — engineers need product sense, product people need engineering/UX sense — and engineers should invest in product, technical evangelism, and go-to-market skills: "Don't be just an engineer" @GergelyOrosz. Firsthand announcement; the source carries topic pointers, not the agent-workflow substance itself. Episode: https://x.com/GergelyOrosz/status/2090157324853711104

Simon Willison (@simonw) reports a firsthand Claude Code for web experiment using smolvm as a code execution sandbox. When Fable 5 spotted that the sandbox couldn't run the task (no /dev/kvm), it didn't ask first — it wrote a GitHub Actions workflow and pushed it directly to GitHub to run the experiments instead @simonw. Notes from the experiment: https://simonwillison.net/2026/Aug/19/smolmachines-untrusted-sandbox/ ; research report: https://github.com/simonw/research/tree/main/smolmachines-untrusted-sandbox#readme@simonw.

Jeremy Morrell hypothesizes a new opportunity for extensible software on the web: LLMs radically lower the cost of authoring extensions, and modern sandbox primitives lower deployment cost while providing security boundaries. His recommended pattern: build an app as a solid, accountable core, then let users safely extend it in many directions by having LLMs fill in the missing pieces. simonwillison

OpenAI's developer account shared a Codex case study in which Asana completed a frontend test migration from Enzyme to React Testing Library in two calendar weeks, versus the five more years originally expected — a dramatic, vendor-reported productivity gain with no workflow details or independent verification @OpenAIDevs.

Kent C. Dodds announced the release of his kody project at v2026.08.19, claiming the release was built in a single day by a solo dev @kentcdodds. The post is firsthand (Dodds owns the repo) but provides no details on what kody is, what changed, or how the work was done — only the link: https://github.com/kentcdodds/kody/releases/tag/v2026.08.19@kentcdodds. The single-day velocity claim is unsubstantiated hype rather than a replicable workflow.

• Simon Willison argues that lines of code can be a meaningful productivity metric for coding agents, despite common wisdom to the contrary. He notes that before agents, a developer producing 200 lines of working, debugged, production-level code was an excellent day, with 50-60 lines typical. With agents, producing 1,000 lines of maintainable, tested debugged code per day is a real improvement — but only achievable with senior-level skill and experience simonwillison.
• The new limiting factor is cognitive capacity: agents let him "churn out code a hundred times faster," but he cannot stay on top of 100x the amount of code, so teams are still needed to load-balance cognitive capacity across engineers simonwillison.
• Coding agents threaten conceptual integrity (from The Mythical Man-Month): features appear in minutes, so software "grows little weird bumps in funny different directions," like the Winchester Mystery House — cheaply added rooms with no overall integrity, making decisions harder simonwillison.
• The practical discipline tip: time cost used to enforce discipline (a crazy feature that took a week wasn't justifiable), but agents make features cost an hour, so developers must consciously impose their own discipline to avoid degrading the codebase simonwillison.

Simon Willison tested smolvm 1.8.3 (smolmachines.com) as a sandbox for untrusted Python and JavaScript data transformations, using hardware-isolated VMs rather than shared-kernel containers simonwillison. Verified working: offline local images, no-network execution, CPU/RAM limits, guest-enforced timeouts, storage quotas, read-only input mounts, writable output mounts, and --unprivileged; cold starts ~0.6–1.5s, warm executions ~50ms simonwillison. He delegated the evaluation to Claude Fable 5 in Claude Code for web with this task: run untrusted Python/JS with RAM and CPU limits (protection against while true), no network access, filesystem access only to designated files, for executing user-provided tasks like data transformations simonwillison. The agent hit an environment limit — the Claude Code web container is itself a Firecracker guest with no /dev/kvm or vmx/svm CPU flags, so smolvm machine run fails with "kvm not available" simonwillison. Its workaround: GitHub Actions ubuntu runners expose /dev/kvm, so the agent ran the real test battery via a temporary workflow on the branch, collected logs, then removed the workflow before committing simonwillison. Replicable pattern: when an agent's own sandbox lacks nested virtualization, offload VM-requiring verification to CI runners that expose /dev/kvm; Willison calls this an example of Fable being "relentlessly proactive" simonwillison. Research notes are at https://github.com/simonw/research/tree/main/smolmachines-untrusted-sandbox#readmesimonwillison.

Firsthand from @kodykoala (Kody's account): Kody v2026.08.19 is out, adding a read-only explorer for published package files — "walk the file tree of any community listing or your own saved package at /files — no git session required" @kodykoala. Full release notes: https://github.com/kentcdodds/kody/releases/tag/v2026.08.19@kodykoala.

Same release, shared by Kent C. Dodds as "Primitive pruning": MCP search and capability discovery is now progressively disclosed — "Discovery returns counted domain indexes for broad queries, ranks your saved packages above raw provider operations, and bounds related-operation detail so agents get what they need without a firehose" (#1533) @kentcdodds.

Breaking changes in this version: package services, values, and remote connectors are retired. Values are absorbed into memories/package storage/repos/secrets/integrations; the Values nav item is gone but /account/values remains reachable for migration; remote connectors are replaced by outbound MCP @kentcdodds.

Kent C. Dodds (@kentcdodds) made a Kody Discord, then wired up Kody to manage it; he reports it's much better telling an agent to do stuff in Discord than trying to click around the UI — a firsthand endorsement of chat/Discord as a control interface for coding agents @kentcdodds.

@jasonzhou1993 announced "OpenRouter, but for agent tools" at https://treg.to, with a video demo, explicitly referencing Patrick Collison's OpenRouter-Stripe news @jasonzhou1993.

Patrick Collison announced that Stripe agreed to acquire OpenRouter — by his account "by far the best way to use new models and manage multiple inference providers" — and argued that "in the future, every business will have to manage both revenue flows and token flows," positioning OpenRouter as "the world's leading token marketplace" @patrickc.

Kent C. Dodds (@kentcdodds) shared a firsthand look at his Devin coding-agent workflow: a parent session titled "Upgrade Remix v3 beta" coordinated five child sessions, each upgrading a separate repository to Remix 3.0.0-beta.10, while the parent session waited for a GitHub API rate-limit reset — an orchestration pattern for fleet-wide upgrades using parallel per-repo child sessions with rate-limit-aware coordination @kentcdodds.

Early-access user @tharshan_09 reports that across different agent harnesses, Kody is the first and basically only MCP they install, but they're unsure when the right task is for Kody; they also built their own package for a hosted agent plans app @tharshan_09. Kent C. Dodds welcomed this feedback from early-access users @kentcdodds.

Kent C. Dodds (@kentcdodds), a prominent developer and educator, reports firsthand that Grok 4.6 has worked so well for him that he no longer uses other models, and he is looking forward to Grok 4.7 @kentcdodds.

Google AI Studio added GitHub support: users can now import GitHub repos, sync bidirectionally (push/pull), and use a new UI for force pushes and merges @OfficialLoganK. Logan Kilpatrick (@OfficialLoganK) said the push/pull feature took a while to land but is now live, with a demo video embedded @OfficialLoganK.

Cursor's official 08-19-26 update @cursor_ai adds several agent-workflow features:

• Cloud agents now pick up work from events, hold a goal until it's met, and stay on course through long sessions @cursor_ai.
• Cloud agents can monitor PRs, watch a Slack thread, or run scheduled tasks; they automatically subscribe to PRs they create and drive them to completion @cursor_ai.
• Any skill can become a Custom Mode (always-on pinned skill): from /, pick a skill and press ⌥⏎ (Mac) or Alt+Enter (Windows), or choose "Use as Mode" @cursor_ai.
• Subagents can now run on their own virtual machines, each with an isolated copy of the project — to test the parent agent's changes in a fresh environment or swarm independent fixes @cursor_ai.
• Steering now waits for the next tool call instead of cutting the agent off mid-action: type a follow-up and hit send, or press ⏎ twice @cursor_ai.
• Use /goal to give the agent a long-lived objective to work toward until fully complete @cursor_ai; full changelog: https://cursor.com/changelog/08-19-26@cursor_ai.

Theo (@theo of t3.gg), creator of T3 Code, is pushing back on people calling T3 Code a "harness," comparing the effort to convincing people that NextJS isn't an alternative to React @theo. @ThadDoormat disputes the claim, saying T3 Code "very clearly is" a harness and asking Theo to explain what a harness is, give examples, and justify why T3 Code is not one @ThadDoormat.

Kody's official account announced that Kent C. Dodds published a new Product Hunt integration for the Kody coding agent (@kentcdodds/producthunt), letting the agent pull today's Product Hunt hunts, hunters, topics, and comments, plus make generic GraphQL requests on the saved OAuth integration @kodykoala; it's hosted at https://kody.codes/@kentcdodds/producthunt@kodykoala. Kent shared the announcement approvingly ("Yo, so that's neat") @kentcdodds. Announcement from the platform, not a hands-on workflow report.

Kent C. Dodds (@kentcdodds) shows a practical Cursor cloud-agent workflow for timed feature removal: with a one-month sunset timeline, he tells the cloud agent to use Kody to schedule itself to wake up when the time is up — the draft prompt (in screenshot) asks the agent to schedule a wakeup near the planned sunset date, review the situation, and recommend next steps before acting @kentcdodds. The setup: Cursor Cloud mode, model 'Grok 4.6 High Fast', branch 'cursor/hide-values-nav-8e48' @kentcdodds. This is a human-in-the-loop pattern where the agent self-schedules a future check-in and is instructed to review before acting — first-hand, from a well-known dev educator.

Kent C. Dodds (firsthand, @kentcdodds) reports creating and merging 41 PRs in a single day @kentcdodds. He adds that this count does not include agents committing to his own private "git hub" inside Kody itself @kentcdodds — a signal of the volume of agent-driven PR work a top practitioner sustains, and that such tallies often undercount internal agent activity.

Kent C. Dodds (@kentcdodds) argues that a viral video being shared as a vibe-coding cautionary tale is mislabeled: 'This is funny. But it has nothing to do with vibe coding and everything to do with product engineering!' He says this failure pattern predates AI coding tools — 'The exact same thing has happened to companies forever' — and pinpoints the root cause: 'It's too easy to focus on the solution and forget to validate the problem is good enough' @kentcdodds. He was responding to a video posted by @kiwicopple @kiwicopple.