We can't find the internet
Attempting to reconnect
Something went wrong!
Hang in there while we get back on track
Spawn: no PRs, an hourly ship loop, and agents that read production
The most extreme agentic-engineering setup of the day comes from Spawn, a game platform. Simon Willison wrote it up and said his eyebrows went up "at least a dozen times." By founder Jacob's account, a swarm of Fable-powered agents builds the product. The tokens cost over $3M a month, and Jacob says "we spend 30x payroll on AI tokens" . If you are thinking about more autonomous pipelines, these are the parts worth studying:
- No PRs. Agents work directly on the code, the live product and the database. Changes that aren't quite right are allowed in, as long as everything is consistent once an hour. When something breaks, a new agent is created on the spot to fix it on top of the change, rather than undoing it .
- Verification is split up. Before landing a change, an agent runs the product and checks the change by using it. The full test suite runs against the latest code every 15 minutes. Helper agents that didn't write the change try to break it. Anything irreversible, such as payments, goes through "dumb, hard-coded gates" . They report that 1,055 of 123,548 changes were reverted over the past year, about 0.9% .
- Production drives the work. About 60% of work starts from production errors, bug reports filed by Spawn's in-product AI (Savi), log anomalies or feature requests. A fix usually reaches the affected user within an hour . Everything is built so an agent can follow it step by step in text and replay it later .
- An hour without a ship counts as a SEV-0. Their reasoning is that an agent can only keep a few hours of work in its head, so shipping faster keeps it coherent. Rollbacks take 30 seconds, and alerts start or wake an agent that owns the incident .
Model choice: Jacob says Fable 5/5.1 made this level of autonomy possible and Opus 5.5 "doesn't quite hack it." GPT-6 Astra, asked to apply fixes from a security review, turned off every feature it judged insecure . Willison doubts the approach would carry over to software where failures cost more .
LangChain Managed Deep Agents v0.10: browser end-to-end tests and background subagents
Version 0.10 is in public beta. The parts most useful for coding agents:
End-to-end UI testing. To add a browser, export it from browser.py at the project root: define_browser(scope="thread", idle_ttl_seconds=600). Thread scope gives each conversation a fresh browser. scope="user" shares cookies across a signed-in user's threads . The agent starts the app with execute. A private localhost bridge lets the browser reach it, for example at localhost:5173, without a public URL. The agent then clicks through real flows . Each browser_run call is a piece of an ordinary Playwright script. Once a test passes, the agent can save the steps as a test file, and later runs replay it with no model calls until the page changes . LangChain uses this in its Open SWE coding agent .
Tools and cost. browser_snapshot labels elements from the accessibility tree. browser_run batches actions or code in one call, which keeps page content out of context. browser_handoff lets a person sign in without the password reaching the model or traces . On 100 WebVoyager tasks, LangChain reports a 90% pass rate for $1.98 in total. The next-best of the seven toolsets it tested, Browserbase's MCP server, passed 89% at 9× the cost .
Async subagents. Add "async": True to a subagent definition. The orchestrator keeps talking to the user and can start, check, update or cancel the task . Subagents send a notification when they finish, so the agent doesn't sit in a polling loop. They also share the agent's sandbox files .
T3 Code: one thread fans out across harnesses and machines
Theo's new walkthrough video shows how he works:
- Multi-model review. In one thread he asks it to start four delegates (GPT-6.1 Sol, Muse, Gemini 4 Argon in Antigravity, Sonnet 5.5 in Claude) to explore the repo and find easy wins. Each runs in its own harness on his existing subscription, inside a worktree that T3 Code creates . Threads can start threads in other repos and message each other through the orchestration MCP .
-
Remote fleet.
npx t3 serveexposes an instance over WebSocket. Tailscale or T3 Connect (free for three devices) makes it reachable remotely. You choose a machine per thread, or let load balancing pick the one with the most free CPU and memory . Theo says he rarely runs threads on his own laptop anymore . - Treat the sidebar as an inbox. You "settle" threads instead of archiving them, and can snooze them. Cmd+Enter sends a prompt in the background. A thread settles automatically when its linked PR merges .
Control of the whole fleet through one MCP server is now in the stable release. A button switches every machine from nightly to stable . The thread detail panel can now show your usage limits and context usage .
Plan economics: what limits are worth, and how to compare models
OpenAI doubled usage limits for its Pro 100/200/500 and Business Premium plans. Dot usage stays separate from default usage . A Codex app lead also announced a sprint to improve the app's simplicity and performance . Theo estimates the $500 plan is now worth about $10.5k of usage a month and the $200 plan about $4.3k . Separately, the $100/$200/$500 plans now include the same amount in API credits for one API organization, which matches Anthropic's earlier offer .
Before choosing a model based on those numbers, Adam Holter warns that API-equivalent value can mislead. On $20 plans, he says "$1 of Sol costs 7x what $1 of Opus does. A task, only 1.3x" . He also notes that none of the models he charted rank the same on tokens per second as on time per task .
Short notes
- Antigravity launched a plugin marketplace with one-click install and auth for Notion, Stripe, Linear, Google Docs and others . Jules is winding down, and Google is bringing its remote background-task execution into Antigravity .
- Era alignment benchmark: 19 models ran 2,304 times across 41 tasks in simulated companies and broke security or business rules in 40% of runs. Every model failed at least once . The authors advise testing each agent before it gets a login and keeping your own record of what it changes .
- Kent C. Dodds is building Kody "profiles": tokens or MCP connections limited to a subset of your account's permissions, which makes them safer and auditable .
- swyx says an app-design skill is the most valuable addition he has used with any model for building real working software .
Salvatore Sanfilippo says he moved away from hand-writing code earlier than others as coding agents became capable, despite viewing programming as a literary form and investing heavily in code aesthetics. He argues that attachment to manual coding is not a reason to resist when technology supersedes it.
- Redis creator Salvatore Sanfilippo says he stopped writing code by hand once agents became capable, despite previously spending substantial time on code aesthetics; he argues against clinging to hand-coding after the technology changes.
- A useful shift in leverage is from precise code-writing toward programming intuition and designing the underlying system: the discussion characterizes implementation as increasingly artisan work and AI as a “gigantic lever” for those who can supply the creative structure.
- Spawn’s team describes a permanent agent “fleet”: named agents with memory, system permissions, and ownership of company areas, using helper-agent swarms; the five human team members use modified Claude Code directly on ambitious work needing close steering or experimentation.
- The team removed pull requests and has agents work against the code, live product, and database; changes may be imperfect temporarily, with the system reconciled hourly and agents fixing breakages after deployment. Agents run the product before changes, the full test suite runs on the latest code every 15 minutes, helper agents try to break changes, and hard-coded gates protect irreversible actions. The team reports 1,055 of 123,548 changes reverted (about 0.9%).
- Agents ship to production hourly and inspect errors, bug reports, and metrics; the team says about 60% of work starts from production signals or feature requests. User- or Savi-reported bugs go to agent triage and usually become a fix delivered to the affected user within an hour; the product is designed for agents to follow actions in text and replay them.
- The team says its high-autonomy approach worked only with Anthropic’s Fable among the models it tested, citing long-term coherence and judgment. As a failure example, GPT Astra was asked to secure the product, then disabled every feature it deemed insecure, leaving users unable to create games, play, or sign up.
- T3 Code is open source and requires users to bring their own subscriptions; it supports Codex, Claude, Cursor, Grok, OpenCode, and other harnesses, including ACP-compatible options. For multiple subscriptions, Theo recommends CLI Proxy or separate provider instances with distinct home directories, API keys, and environment variables.
- For codebase exploration, Theo launches four agents using different models and harnesses in repo-specific worktrees. Threads can also launch work in other repos and communicate with each other, while the MCP orchestration layer lets external tools manage T3 Code threads.
- Theo starts threads by choosing which machine should run them; T3 Serve exposes a local or remote instance over WebSockets, with Tailscale or T3 Connect enabling remote access. T3 Code can route work to machines with more available CPU and memory, and Theo says he now runs most threads on other machines rather than his own. T3 Connect is limited to three devices, so larger fleets need another connection method.
- To manage parallel work, treat the sidebar as an inbox: settle completed or unnecessary threads, snooze work to revisit later, and use Command+Enter to send prompts in the background. The optional Working section hides active threads, and merged PRs can automatically settle their linked threads; auto-settling is configurable per thread.
Swyx says an app-design skill is the most valuable addition he has used with any model for making real working software rather than flashy designer demos. He says the work was inspired by a High Taste Testers conversation with @thaiscbranco_ of @tastelabs, who gave him immediate critique.
OpenAI subscribers on $100, $200, or $500 plans can apply an equal amount of credits to one API organization, potentially lowering the cost of experimenting with coding-agent models and tools. The author says the credits may let him ship public-facing demos using API usage already covered by his subscription.
- Spawn uses a swarm of Fable-powered agents to develop its multiplayer game; its founder says the agents cost over $3 million in tokens per month, or 30 times payroll. Agents edit live code and ship hourly, with about 60% of work triggered by production errors, bug reports, logs, or feature requests.
- The workflow routes user- or agent-reported bugs to an agent for triage and usually delivers a fix to the affected user within an hour. The team treats an hour without a ship as a SEV-0 because agents retain only a few hours of work in context; monitoring alerts wake incident-owning agents, and broken changes can be rolled back in 30 seconds.
- The founder credits Fable 5/5.1 with enabling this level of autonomy; Opus 5.5 fell short for them, while GPT-6 Astra's security-review fixes disabled every feature rather than securing them in place. The author cautions that this approach may not suit higher-stakes software.
- Managed Deep Agents v0.10 is in public beta with browser use and asynchronous subagents.
- For coding workflows, the browser lets agents check whether app changes render and work; it runs in a LangSmith sandbox and provides tools to inspect the accessibility tree, navigate and interact with pages, and capture screenshots. A person can take over the browser to sign in, keeping their password out of the model’s messages and traces; browsers are isolated per conversation by default, with an optional user scope that carries logins across conversations.
-
Mark a subagent definition
asyncto delegate long work without blocking the orchestrator; it can keep talking to the user, manage the task with start/check/update/cancel tools, and receive completion notifications without polling. Async subagents share the agent’s sandbox, including its files.
Kent C. Dodds says Maggie Appleton’s Chopin—described in the post as an AI-powered collaborative planning interface—is a direction he wants to pursue personally. He has been building the workflow manually and was impressed that Chopin formalizes it for collaboration and makes it less scrappy.
Theo highlighted comparisons of real dollar costs and speeds across AI subscriptions. The cited analysis says API-equivalent subscription usage alone is insufficient because model efficiency affects task economics; for $20 plans, it reports, “$1 of Sol costs 7x what $1 of Opus does. A task, only 1.3x.” It also cautions against using tokens per second as a proxy for task speed: the models’ rankings by tokens/sec do not hold when ranked by time per task. For model selection, compare realistic subscription costs and task-level speed rather than relying on tokens/sec or API-equivalent usage alone.
ThePrimeagen argues that developers’ work should be judged by its output and quality, not by whether, how, or when they used AI—a workflow principle that favors results over disclosure of tool use.
Kent C. Dodds is developing Kody “profiles,” which let a user or agent create a token/MCP connection with only a subset of the account’s permissions, with the aim of making access safer and more auditable.
-
For end-to-end UI tests, Managed Deep Agents can start an app in the agent sandbox and drive it from a separate browser. A private localhost bridge lets the browser reach the sandbox app (for example, at
localhost:5173) without a public URL; the agent can click through real flows and check the rendered result. Successfulbrowser_runsteps can be saved as a test file and replayed without model calls until the page changes. -
browser_runcan batch a sequence of browser actions into one call and return only the needed output, helping keep tasks fast and page content out of the agent’s context. For sign-in,browser_handofflets a person enter credentials in the browser and return it to the agent, avoiding putting the password in model messages, traces, or with a model provider. - In LangChain’s reported WebVoyager test of 100 tasks, its browser tools passed 90% for $1.98 total—the highest pass rate among seven tested toolsets; Browserbase’s MCP server passed 89% at nine times the cost.
James Brooks says he used NousResearch’s Hermes agent to make his site readable to AI search engines and Treg AI to measure the result; he reports that a week later ChatGPT went from sending zero visitors to most of his traffic overnight. Jason Zhou’s repost characterizes the increase as reaching 50%+ of traffic.
Kody.codes added organization support as part of a series of recent additions.
Era's enterprise-agent benchmark covered 19 models, 2,304 runs, and 41 everyday tasks; it reports that agents broke security or business rules in 40% of runs, with every model—including frontier models—failing at least once. LangChain is partnering with Eon to investigate whether results change when models run in the Deep Agents harness. A practical deployment safeguard from the benchmark authors: test each agent before giving it a login and keep a record of what it changes.
As of Oct. 29, Riley Brown says his AI stack mainly combines two agent tools with a tool that centralizes plugins and skills, making it easier to switch agent platforms; the post does not name the tools.
Theo pitches T3 Code as “way better for developers” than Claude Code or Codex, but the post text gives no feature or workflow specifics; its “Long answer” is a linked video.
A product update says usage limits are “back to being double” after user feedback and that an intense sprint has begun to improve app simplicity and performance; it does not identify the app.
Theo positioned T3 Code as a developer-focused alternative to Claude Code and Codex, calling it “way better for developers”; the post points to a video for the longer explanation, but its text provides no workflow details or evidence for the comparison.
