Why Simpler AI Coding Workflows Win

Theo’s workflow has completely inverted from his prior advice. Where he once recommended elaborate plan modes and parallel work trees, he now advocates for radical simplicity: one thread, one task, minimal prompting. His core insight is that engineering teams over-complicate their agent interactions. Developers obsess over file paths, skills, and bloated agent.md files, but the real leverage comes from reading the model’s text carefully and steering it conversationally. The model is smart enough to find the right files and infer intent if you stop micromanaging it.

His toolchain reflects this philosophy. He’s abandoned Claude for GPT-5.5 on the Codex harness, favoring its boring CLI but rich desktop app—especially for remote execution. The inability to paste images or maintain stable sessions over SSH makes traditional terminals a non-starter for serious agentic work. Remote coding via a browser-based UI, often on a separate machine, decouples agent execution from his local laptop, allowing him to close his lid without killing long-running tasks. T3 Code’s remote hosting shines here, though Codex’s mobile integration is promising.

His project, Lakebed, was built almost entirely through single-shot prompts: two-sentence requests that produced working implementations in under 10 minutes. He rarely branches or uses work trees; old context is discarded aggressively to prevent the model from drifting. When the agent pushes back, he listens—scope is negotiated, not dictated. Verification is crucial: he gives agents CLI commands, test suites, or computer-use tools to self-validate, reducing the back-and-forth on broken PRs. For complex changes, he offloads the grunt work to Code Rabbit, letting the agent loop until all review issues resolve.

The hidden lesson is about context hygiene. Every token in chat history biases the model, so treating threads as disposable, single-concern containers produces far more reliable outputs than elaborate persistent sessions. This discipline, combined with a lightweight agent.md that states high-level intent rather than technical specs, makes the workflow fast and predictable. For engineering leaders, this suggests a counterintuitive playbook: invest less in prompt engineering and more in a workflow that ruthlessly simplifies context and maximizes output readability.