How Chain-of-Thought Turns Reasoning Models Into Constraint Solvers

▶ Watch (00:44)

Reasoning models solve problems “an order of magnitude more complex than what we could do with general LLMs,” because they run chain-of-thought internally before answering. Older models answered a duck-eggs math problem wrong when prompted directly, then got it right with the exact same parameters after few-shot examples forced step-by-step logic. New reasoning models do that deliberation automatically.

Azure AI Foundry’s catalog lists options from Microsoft, OpenAI, and DeepSeek, including Phi-4 with reasoning as a small-model choice. The visible “think” block in GitHub Models shows the model organizing its answer before output. The tradeoff: every internal reasoning step burns tokens and adds latency.

The Summer Camp Scheduling Demo: Reasoning Models Against Real Complexity

▶ Watch (09:18)

Marsman pasted a 12-page summer camp brochure as raw unformatted text into o4-mini via Azure AI Foundry, then asked it to schedule three kids, each at a different grade level, for the full summer with two family weeks off and no repeated camps. In the o4-mini summer camp scheduling demo (11:45), every camp verified against the source: “Fashion Design” matched July 8th, “Baker’s Workshop” the 22nd, and the 19th correctly showed nothing. No hallucinations.

The non-reasoning model put a 12th-grader in “Song Studio,” a class the brochure lists for 7th and 9th graders. Same prompt, same data, wrong grade levels.

Picking the Right Model: When Reasoning Hurts More Than It Helps

▶ Watch (15:43)

A simple story problem with a correct answer of 1,100 ran fine on both GPT-4.1-mini and a reasoning model. GPT-4.1-mini finished faster. The reasoning model spent tokens deliberating on a question that didn’t need it. The takeaway: “don’t use a reasoning model if you don’t need one.”

For teams that can’t predict query complexity, Microsoft announced Model Router at the conference. It dynamically routes each request to a reasoning or non-reasoning model based on the query’s difficulty. Marsman shared a public repo (JenniferMarsman/o1Analysis) with all four benchmark problems, including scheduling, document comparison, Minecraft function-calling, and calendar conflict checks.

Computer Use Agents: Vision-Only vs DOM-Aware and What That Means in Practice

▶ Watch (25:58)

Marco Casalaina contrasted two CUA types. The Responses API model operates a full Linux VM through computer vision alone: it sends screenshots with every click. On a date picker it clicked a few pixels off and started typing in the year field before the month field.

“Date pickers are the bane of Computer Use agents.” — Marco Casalaina

In the Browser Use agent shopping on Amazon (31:05), DOM-based click targets replaced pixel guessing. Browser Use left Nordstrom when it stalled and switched to Amazon, illustrating the non-determinism all agents share. Casalaina flagged LLMs.txt as an early standard for exposing agent-friendly APIs alongside traditional web pages.

RPA and CUA as Partners: UiPath’s Micro-Task Approach to Enterprise Automation

▶ Watch (40:29)

Cosmin Voicu argued CUAs hit acceptable reliability only on micro-tasks, sequences of a few clicks. Longer tasks drift and cost too much. The fix is chaining micro-tasks with RPA and API calls on an enterprise platform. In the UiPath UI agent replacing RPA workflow with a prompt (50:35), a full RPA block for table extraction, filtering, and dynamic clicks collapsed into one natural-language prompt.

The UiPath Fortnite build compatibility testing demo (52:52) showed what RPA cannot handle: hundreds of game builds across different languages and UI layouts. The agent clicked the right Play button on each.

“It’s not super-fast, and it’s not super-cheap,” — Cosmin Voicu

Q&A

What are the key differences in building these workflows in UiPath versus Copilot Studio? Casalaina said Copilot Studio is more of an agent designer and that Power Automate is a closer analogy; Voicu noted UiPath lets you mix RPA and agents directly, and both agreed the capabilities are broadly similar with Copilot Studio likely catching up. ▶ 55:29

Notable Quotes

Date pickers are the bane of Computer Use agents. Marco Casalaina · ▶ 28:45

Like, reasoning models absolutely shine Jennifer Marsman · ▶ 20:26

It’s not super-fast, and it’s not super-cheap, Cosmin Voicu · ▶ 54:22

Key Takeaways

  • Reasoning models solve constraint-heavy problems by doing chain-of-thought internally, but cost more time and tokens than standard models.
  • Vision-only Computer Use Agents make pixel-click errors on dense UI; DOM-aware agents like Browser Use land clicks more reliably.
  • Use CUAs for micro-tasks requiring visual reasoning, RPA for everything deterministic, and combine them on an enterprise platform.