So my workflow has changed a lot...

Summary

The stream covers a major change in Theo’s development workflow: instead of personally tracking every coding task, he uses AI agents, reusable skills, and T3 Code’s inbox-style interface to delegate more work and review it at key decision points. He also previews a Samsung foldable as a way to assess the iPhone Duo form factor, gives a broad T3 Code walkthrough, and analyzes Anthropic’s new internal source-control system.

Foldable-phone impressions

Theo says the Samsung Z Fold 8 is a surprisingly good preview of the form factor he expects from the iPhone Duo. He likes its aspect ratio, one-handed size, and tablet-like experience when rotated for scrolling or coding. He reports that the crease is barely visible in person, though lighting makes it show up on camera.

His main complaint is the speakers: in landscape orientation, they can sound unbalanced or effectively mono. He also adjusted the keyboard by disabling its suggestions and number bars to free up screen space.

AI-agent skills and workflow

Theo explains three Markdown-based skills that encode work patterns he developed through repeated use:

• Soul Loop: Uses a separate model, described as GPT-6.1 Soul, to review code changes in a bounded loop. It aims for the smallest change that satisfies the intent, avoids scope creep, and allows the agent to push back on unhelpful review comments. Theo’s version limits the review to three rounds by default, with more rounds reserved for serious risks.
• Full Send: Gives an agent permission to take a task from its current state through verification, pull-request review, and merge. The skill requires evidence that the change works, passing checks, and resolved or explicitly dismissed review findings. The agent should ask Theo for help only when blocked or unable to proceed safely.
• Full Stop: Uses a similar autonomous workflow, but stops when the pull request is ready for Theo to merge. This is for changes he wants agents to finish and verify without giving them merge authority.

In a recent sample, these skills appeared in nearly half of Theo’s T3 Code threads. In one sample, 27 of 32 threads using Full Send merged without him needing to inspect them again. He emphasizes that this relies on verification, monitoring, and the ability to catch or roll back problems—not on assuming agents are infallible.

Other supporting skills include taking over an existing contribution while preserving the original contributor’s credit, modernizing or rebasing pull requests, and monitoring PRs. Theo keeps skills in a shared “fleet” repository so he can sync them across machines. He prefers natural-language trigger phrases in skill descriptions over slash commands, since he wants to invoke multiple skills together without remembering special syntax.

The guide and demo cover how to build and revise agent skills from recurring workflow patterns, delegate review to different models, and use a CLI proxy to make models available across tools. Theo cautions that Anthropic is more restrictive than OpenAI about using subscription access through other tools, so users should follow each provider’s terms.

T3 Code walkthrough and product features

Theo presents T3 Code as an open-source environment for working with AI coding agents. Users bring their own subscriptions and compute; T3 Code does not sell access to models. He says the product had grown to roughly 550,000 users and 140,000 weekly active users, according to the figures he cites on stream.

Highlighted capabilities include:

• Multiple agents and harnesses: Support for tools such as Claude Code, Codex, Cursor, Grok, OpenCode, Anti-Gravity, Muse Code, and Pi, as well as other ACP-compatible tools. Users can run agents from different providers in parallel and have them review work or communicate through T3 Code’s orchestration layer.
• Orchestrator V2: Lets agents launch other agents or threads, including threads in different repositories and on different machines. T3 Code exposes these capabilities through an MCP server, so external tools can manage threads too.
• Remote development: Threads can run on selected computers, including remote Linux machines, and be controlled through the desktop app, web app, or mobile app. Connections can use T3 Connect, TailScale, or other routes. T3 Connect is limited to three devices at the time of the stream.
• Inbox-style task management: The sidebar is designed to show work that needs attention. Users can settle, snooze, or archive threads; active work can be hidden in an optional “working” section; and linked PRs can automatically settle their threads after merging.
• Parallel work and PR tools: Users can start work in the background, monitor pull requests within the app, and use built-in watchers rather than asking agents to poll GitHub inefficiently. T3 Code also supports scheduled tasks, webhooks, and switching harnesses mid-thread.
• Development and diagnostics: Features include remote browser and iOS/Android emulator use, machine diagnostics and load balancing, interactive HTML output, image attachments accessible to agents as files, snapshots for capturing context, usage views, and support for multiple Git providers.

Theo notes several trade-offs: users must provide their own machine or cloud compute; remote access can require additional setup; computer use is not yet built in as a general T3 Code feature; and the flexible, parallel workflow can burn through model usage quickly. He also says the nightly build contains major recent changes to orchestration, storage, and connections, while a stable release was being prepared.

Anthropic’s Scion source-control system

Theo reviews an Anthropic article about Scion, an in-house version-control system, and Antique, a code-review surface. Anthropic says its rate of merged changes grew from about 2,500 per day in May to more than 30,000 per day in September, making developer infrastructure a bottleneck. The company says a small team built Scion and related systems and migrated its monorepo in five months.

The main design differences Theo highlights are:

• Online, server-centered operation: Unlike Git’s decentralized model, Scion assumes most clients are agents or jobs that are online. The server performs operations such as commits, merges, rebases, and history queries.
• Smaller, selective checkouts: Clients can request only the files or data they need. Scion also supports virtualized checkouts that materialize files on demand, reducing unnecessary work for CI and agents.
• Content-addressed, immutable data: Changes are stored as change sets that refer to file contents in blob storage. Additional structures for history, blame, and search can be computed as derived data.
• Scalable caching and APIs: The architecture uses separate server and cache tiers, with source-control queries exposed through APIs. Anthropic says the new infrastructure uses far fewer CPU and memory resources than its previous Git-mirroring setup.
• Different performance profile: The article reports large speedups for common operations, including much faster status checks, checkouts, worktree creation, and pulls.

Theo sees Scion as an example of rebuilding tools around agent-scale development rather than treating existing conventions as untouchable. He connects its design to his own emphasis on making merges safer through strong verification, limited blast radius, and quick rollback. He also notes that Scion is highly bespoke and not presented by Anthropic as a general-purpose Git replacement.

As a rough point of comparison, Theo estimates that Git operations and checkouts have consumed around 1,000 hours of T3 Code CI time, though he describes this as an estimate.

Other technical and product notes

• Theo discusses an experimental Rust subset intended to compile much faster than Rust, describing it as a speculative project motivated by slow compile times in his TypeScript-to-Rust work.
• He describes automating post-stream media handling: transferring footage to a server, syncing it to editors, and re-encoding large camera files to H.265 for storage.
• He previews MegaDev, a forthcoming course about using AI in real-world software development. He says it is aimed at people already working as developers, ideally with employer funding, rather than people buying a course to get their first job.
• He argues that model benchmarks and rankings should be interpreted carefully: his personal tier lists reflect how much he likes using a model, not simply its raw capability. He praises Muse’s interaction quality and criticizes Gemini’s coding reliability based on his experience.

Main speakers and sources

• Theo (Theo – t3.gg): Host, product developer, and commentator throughout the stream.
• Anthropic’s blog post on Scion and Antique: The primary source for the source-control architecture and performance claims discussed.
• T3 Code team and community contributors: Referenced in the product walkthrough and acknowledgements, including Julius, Maria, Shiv, Wout, and Bal.