I've been thinking a lot lately...

Summary

The livestream centers on four videos Theo records and discusses: a perspective on AI-era software engineering, a review of Anthropic’s code-review process, a guide to his multi-model “Council” workflow, and a reaction to PewDiePie’s local-AI project.

1. Engineers as “gardeners” in the AI era

Theo argues that software engineers’ work is shifting from writing every line of code to maintaining the environment in which code—often produced by agents—can be built safely and effectively. He calls this “gardening”: improving the codebase, tools, checks, and feedback loops so good changes are easier to make and bad changes are caught early.

• Drawing on his time at Twitch, he describes how product and platform teams need to improve together. Better platforms reduce friction for product work, while product work reveals what the platform needs to support.
• He connects this idea to the T3 Stack and his later move of T3 Chat from a custom synchronization system to Convex. The migration was a major internal overhaul, but it made the backend easier for teammates to understand and work on.
• AI agents make platform maintenance more important: slow startup, CI, or release processes can block many parallel agents at once.
• Examples of “gardening” include removing low-value tests, adding checks for performance regressions, and investigating the underlying cause when agents repeatedly make the same mistake. Theo says one cleanup removed about 258,000 lines of tests; only a few deleted tests were later restored.
• He presents code quality as more than line-by-line review. The goal is to make the whole system easier to change, verify, and recover when something goes wrong. He also emphasizes fast rollbacks and limiting the impact of failures.

2. Anthropic’s approach to code review

Theo reviews an Anthropic article about how the Claude Code team handles a rapidly growing volume of agent-generated changes. The article describes layered checks intended to reduce reliance on manual review of every line.

The process described in the article:

1. Have the authoring agent verify its work before opening a PR—for example, by running tests or checking the app.
2. Run CI, including linting, type checks, tests, and builds.
3. Use automated bug and security reviews, with findings filtered and ranked before they reach people.
4. Match human approval requirements to the change’s risk. Some changes may need no human approval, while sensitive areas require more review.
5. Use merge queues, staged releases, feature flags, monitoring, and rollback procedures to catch problems quickly.

The article reports that Anthropic’s monorepo changes increased from about 2,500 per day in May to more than 30,000 in September. It also reports that merged Claude Code PRs more than doubled from June to September while the share needing reverts fell by over 40%. After changes intended to reduce review loops, more PRs reportedly passed after one review-and-fix round, while fewer required five or more rounds.

Theo’s analysis emphasizes that human review is not a reliable way to catch every bug, and that agents should review and test their own work before asking teammates to look. He favors fixing recurring issues in the codebase or static checks before adding more instructions to agent guidance files. He also warns that noisy feedback can cause agents to add unnecessary code, so review loops need limits and should converge.

3. Multi-model review with the “Council” workflow

Theo explains a workflow for getting feedback from multiple model families on difficult changes. Rather than routing every task to a supposedly ideal model, he uses different models as reviewers for work produced by another agent.

• His Council skill can call models through different coding tools or harnesses, ask them to review a change, and iterate on the feedback.
• The models can run code and test it, rather than only commenting on a diff.
• Theo uses scores to set a stopping threshold, such as requiring reviewers to rate a change highly before it is ready for his attention. He describes related workflows including “Full Stop,” “Full Send,” and “Soul Loop.”
• He gives examples where review caught deleted tests that still protected against bugs, security problems in shared-thread functionality, and edge cases involving interrupted or failed agent runs.
• He says his tracked reviewer scores generally improved over review rounds, while cautioning that this does not make the process infallible.

The main drawbacks are cost and the risk of endless iteration. Theo says the workflow can consume substantial subscription usage and advises against using it at ordinary API prices. He also discusses adding round caps and scope limits because reviewers can otherwise keep generating fixes and expanding a PR without making meaningful progress.

His broader recommendation is to let agents verify and refine work in the background, then bring it to a person when it is ready for a decision.

4. Reaction to PewDiePie’s local-AI project

Theo reacts to PewDiePie’s video about Odysius and its fine-tuned local model, Ajax. In the video, PewDiePie describes a model intended for tasks such as web browsing, email, and calendar use, along with efforts to collect user-submitted data and train the model further.

• PewDiePie says Ajax was fine-tuned for a particular harness and that he used the open-source Heretic project to reduce refusals. He describes training experiments with GRPO, repeating tasks multiple times and reinforcing successful attempts.
• Theo notes that reducing refusals can have broader and unpredictable effects on a model’s behavior, and cautions that heavily “uncensored” models can pose security risks when run locally.
• He discusses why advanced models’ ability to use subagents is difficult to reproduce: useful orchestration data must be created, and some model providers conceal or encrypt parts of that process.
• Theo is interested in local AI but pushes back on the idea that large frontier models are necessarily wasteful for simple tasks, noting that mixture-of-experts systems may activate only part of a model for a given request.
• He also stresses that developers still benefit from understanding code and engineering fundamentals, even when agents write much of it.

Other product and engineering notes

• Theo says T3 Code is not currently in acquisition talks.
• He discusses plans for a T3 Code plugin system, but says the required architecture—especially connections across client, server, machine, and source-control components—is still being worked out.
• He mentions features including T3 Code triage and guest sharing for threads, and says remote control is planned.
• He discusses resource use when running many subagents. In his own profiling, Claude Code and Codex appeared to use more CPU and RAM than T3 Code, though he acknowledges that T3 Code could improve its storage and resource handling.
• He describes using a separate assistant workflow for non-coding tasks such as email, calendar, reminders, and business administration, but says it is too connected to private information to demonstrate live.

Reviews, guides, and tutorials covered

• AI-era engineering guide: Treat codebases and development workflows as gardens; reduce friction, catch recurring problems at their source, and make changes easy to verify and undo.
• Code-review workflow review: A discussion of Anthropic’s layered review system, including pre-PR verification, automated reviews, risk-based human involvement, and staged releases.
• Multi-model review guide: How Theo uses the Council skill to get diverse model feedback, verify changes, set quality thresholds, and avoid unbounded review loops.
• PewDiePie project reaction: An analysis of the Ajax/Odysius local-model effort, its training approach, and the safety considerations of reducing refusals.

Main speakers and sources

• Theo (“Theo - t3.gg”) — Livestream host, software developer, and commentator.
• Anthropic’s Claude Code team — Source of the code-review article Theo analyzes, including material attributed to Addy and colleagues.
• PewDiePie — Creator of the video about the Odysius project and Ajax model.
• Other referenced sources — Lauren/Pstack, Boris of Claude Code, and comments or suggestions from Theo’s livestream chat.