How Thariq Shihipar Plans, Builds and Reviews Long-Running Agent Work

Anthropic's Thariq Shihipar discusses agent loops, planning and video workflows with Peter Yang. He demonstrates a small captioned-video pipeline and explains the exploration that preceded it. The episode is sponsored by Wispr Flow.

Thariq describes loops as ways to organize feedback and longer-running work. He distinguishes a goal with an exit condition from workflows that can split work and verification among agents. The choice depends on the task: a measurable latency target has a different feedback signal from a subjective design review.

For the demonstration, he records a short clip and asks the agent to transcribe it with Whisper, create word-highlighted captions and overlays with Remotion, and render the result. A first output appears from one prompt, but his name is transcribed incorrectly and the overlay placement still needs work.

Thariq stresses that the prompt was preceded by exploration. He asks for an HTML explainer on transcription, including possible failure modes such as hallucinated words during silence or split words. Learning those limitations helps him decide how to design the workflow and review its output.

He also generates visual alternatives for captions and overlays from a style reference. Seeing options helps him decide what he wants. He treats prototypes and readable explanations as planning tools and warns that a long plan has little value if nobody reads the important parts.

Planning continues through implementation: start with a human request, explore, refine mockups or explanations, and ask the agent to record unexpected implementation findings. Those findings can lead to a changed specification. A small prototype can test a direction before a more expensive implementation.

Thariq describes using Claude in Slack for background and collaborative work while focusing on one main task in Claude Code. He says different people use different combinations, and notes that treating an agent as an employee is a useful but incomplete metaphor.

Yang shows his podcast-production skills, which generate text, thumbnails and short clips. Thariq suggests that reusable scripts and a persistent workspace can complement a skill, so the agent does not have to reconstruct every tool from scratch.

For multiple shorts, Thariq proposes a coordinator that selects clips, separate agents that produce them, and a rubric for verification. He says a separate reviewer may be less lenient than a model checking its own output, but the workflow uses more compute and does not guarantee unbiased evaluation.

To manage attention, he tries to concentrate on one primary project rather than running many threads without focus. He notes that a rushed instruction can cost time later, and the right amount of parallel work depends on the person and task.

Thariq says the team shortened Claude Code's system prompt by 80%. His view is that newer models may need fewer examples and that explaining the reason for a preference can work better than an unnecessarily absolute instruction. He still recognizes that genuine hard constraints can be useful.

Finally, he encourages learning enough about a system to identify unknowns and tradeoffs. Rather than merely asking an agent to produce a result, he uses explainers to understand what is possible and where tools can fail. His stated goal is to become better as well as faster.