Dynamic Workflows Emerge as New Paradigm for Test-Time Compute in AI Agents

Recording ready if you would like to learn more: https://academy.dair.ai/events/cmpzo9kk3000g04lgsgf5io3z

@omarsar0 "applies to a very small set of use cases" Finally some clarity. I keep seeing them recommended and struggle to see why outside the narrow, albeit powerful, use cases you mention.

@jgilbertson47 My rule for when to use it is when there are many directions I can take (experiments, stack choices, optimization routes), and I might benefit from different perspectives. But this often happens when the project is a bit fleshed out already.

@omarsar0 the data pipeline used to be hand-tuned, now an agent decides what data the next model needs

@omarsar0 Trace planning boundaries.

this matches what i’m seeing while building mac control

most tasks don’t need a complex workflow. one obvious path is enough

the dynamic part becomes useful when the machine is in an unexpected state: wrong window, blocked page, permission prompt, app not focused

at that point the verifier is doing more work than the planner

@omarsar0 TTC is just another buzzword for heavy prompt engineering. Same old cycle.

@omarsar0 planning >> iteration speed.

one round of asking the right questions saves 10 rounds of wrong ones. most teams treat planning like overhead. it's the product.

@omarsar0 Dynamic workflows offer a strong test-time compute alternative, but orchestration still requires rigid verifiers.

@omarsar0 The benchmarks point is the big one for me. A lot of this sounds right, but it is still hard to measure cleanly today.

@omarsar0 Yes, this mostly reads to me as TTC + stronger verification. The key benchmark is when orchestration beats simply increasing reasoning level on a frontier model.

Your notes basically rediscover the harness taxonomy. Here are the six shapes Claude composes (from Anthropic's "A harness for every task"): verifiers/judges → adversarial verification + tournament; llm council → fan-out-and-synthesize; /loop → loop-until-done. The vocabulary is what lets you nudge it.

@omarsar0 Hill-climbing framing is the key insight; most people conflate TTC with brute-force scaling, but structured search over the solution space is where the real gains live for exploratory work.