Channel: Theo - t3․gg
I guess we're writing loops now?
The Signal
Agentic coding has reached a turning point where human oversight is shifting from granular, step-by-step prompting to designing self-sustaining, autonomous loops. The subject, an active developer, argues that agents should handle the full lifecycle of a task—from PR creation and verification to comment monitoring and re-review cycles—without constant human handholding. The core dispute centers on whether this automation is a reliable productivity multiplier or an expensive way to multiply errors, with the speaker advocating for dynamic workflows over rigid, persona-based subagents.
The Case
- The speaker demonstrates 'looping' where one task autonomously spawns follow-up threads, PR creation, and review cycles; they successfully had four stacked PRs reviewed and merged while sleeping.
- A single, simple code-review comment can trigger massive consumption, with one agent spending over eight hours and 3 million tokens to resolve three minor issues.
- The speaker rejects static 'persona' definitions (like 'security reviewer') as ineffective, arguing that agents must build context dynamically as the specific problem evolves.
- Architecture-heavy tasks can be decomposed by agents; for one project, an agent reframed a single, monolithic refactor into three separate, stacked PRs to improve deployment efficiency.
- The speaker reports high inference costs—averaging $10,000 monthly across their machines—which they justify as a worthwhile expense since they remain within their high-tier subscription plan limits.
- The speaker warns against using autonomous loops on mission-critical, large-scale production systems, stating that the current reliability is insufficient for codebases with millions of users.
The 1 Minute Signal Take
The evidence supports the speaker’s claim that current agents are capable of handling multi-stage, autonomous loops, but their economic and reliability arguments are strictly anecdotal. The video is worth watching if you want to see the specific 'heartbeat' workflow patterns currently used to chain PR threads, as this visual context is far more useful than the speaker's repetitive, overconfident generalizations on general agent productivity.
Time saved:
Tags
Channel: Theo - t3․gg