Copilot Studio: Build Custom Copilots

Key insights

• Copilot Studio: The video explains the redesigned Copilot Studio, a low-code, natural-language-first tool for building AI agents. It runs new agents on the GitHub Copilot harness and offers a unified authoring surface for quick creation and deeper customization.
  
• harness: A harness is the runtime layer that steers agent behavior separate from the underlying model. The video compares three harnesses—Copilot Chat, Standard, and GitHub Copilot—and notes you cannot switch a harness after you create an agent.
  
• workflow designer: The new authoring UI includes Build, Preview, Evaluate, and Monitor tabs plus a workflow designer. You can add AI nodes like Agent, Classify, and M365 Copilot alongside connectors, conditions, and loops to assemble flows visually.
  
• skills: The studio replaces manual "topics" with plain‑language skills you write to describe agent behavior. The platform follows a simple lifecycle: Create, Build, Test, Publish, and Monitor to move agents to production.
  
• credits billing: The video warns billing starts during the build phase—each test chat in Preview consumes credits. Also note the pricing deadline: agents created before August 3 keep old pricing only until September 1, so plan migration and budgets accordingly.
  
• capabilities: Copilot Studio supports real‑time voice agents, app and data integration, agent quality evaluation, and multiple model choices (including newer family models). The video advises choosing the right harness and setting limits to avoid surprise costs while testing.
  

Anders Jensen [MVP] published a detailed YouTube walkthrough explaining the August 3 update to the new Copilot Studio, and this article summarizes his key findings for newsroom readers. In the video, Jensen demonstrates how every new agent now runs on the GitHub Copilot Harness, and he walks viewers through the redesign of the authoring surface and the runtime model. Consequently, the update shifts the platform toward a natural-language-first workflow, and it aims to make agent authoring faster for makers who want lower-code experiences.

What the Update Changes

First, Jensen explains the difference between a model and a harness, clarifying that the harness orchestrates how models and skills work together at runtime. Moreover, the new experience replaces old constructs like explicit topics with skills written in plain language, so creators describe intent and the system generates much of the underlying flow. As a result, Microsoft 365 positions this as a production-ready preview that unifies authoring, testing, and Monitoring into one surface.

Second, the runtime shift matters because not all harnesses behave the same way, and Jensen compares three harness options: Copilot Chat, Standard, and GitHub Copilot. He highlights that the harness choice is permanent, and therefore creators must decide up front whether they want the specialized behavior of the GitHub harness or the broader options of the standard harness. Finally, the rollout also affects existing agents: those created before August 3 keep their prior pricing only until September 1, which makes timing critical for Teams that must manage budgets.

Inside the New Studio

Jensen tours the authoring tabs—Build, Preview, Evaluate, and Monitor—showing how the interface encourages writing plain-language instructions to create skills. Additionally, the new workflow designer places AI-focused nodes like Agent, Classify, and M365 Copilot alongside classic connectors, conditions, and loops, which helps bridge AI-driven steps and traditional automation. Therefore, the platform supports multimodal agents and richer orchestration without forcing authors to handcraft every branch.

However, Jensen also demonstrates a cost-related surprise: billing starts while you build, because every test chat in Preview consumes credits. Consequently, Teams that run many interactive tests can find costs accumulating quickly, and Jensen warns that creators should set limits before exploratory testing consumes a budget. He further notes that the studio supports evaluation tools and Monitoring, which helps detect quality regressions but does not remove the need for careful test planning.

Tradeoffs and Design Choices

The update trades detailed manual control for ease and speed, and Jensen frames that tradeoff clearly: natural-language-first authoring lowers the entry barrier, yet it reduces direct control over explicit conversational structures. On one hand, makers gain faster iteration and simpler authoring; on the other hand, advanced teams that need fine-grained flow control might find the abstraction limiting. Therefore, organizations should weigh whether the convenience of auto-generated flows outweighs the need for predictable, hand-tuned behavior.

Moreover, the irrevocable harness choice introduces a longer-term planning challenge because you cannot convert an agent between the GitHub Copilot harness and the standard harness later. Thus, teams must align harness selection with both technical goals and compliance or model-preference requirements. Finally, billing during preview forces a tradeoff between rapid interactive testing and budget discipline, so operational controls and governance are essential.

Practical Challenges and Risks

Jensen points out practical risks around cost, migration, and observability, starting with preview credits that can be spent unexpectedly if tests are frequent or automated. Consequently, project leads should configure limits, track consumption, and coordinate test plans so a single exploratory session does not use a disproportionate share of credits. In addition, the inability to switch harnesses means migrations require reauthoring or rebuilding agents, which raises both time and technical debt.

Furthermore, the move from topics to skills implies new testing strategies, because plain-language skills may produce broader behavior than tightly scoped topics did previously. Therefore, teams must invest in evaluation steps and monitor agent outputs in production to catch unintended responses. Jensen’s walkthrough also underlines the importance of choosing the right connectors and building robust test suites to validate integration points and data access.

Advice for Teams and Next Steps

Based on Jensen’s video, the immediate action is to review any agents created before August 3 and confirm whether you need to preserve old pricing before the September 1 cutoff. Next, when designing new agents, identify your priorities—model behavior, cost controls, or integration complexity—and choose the harness that best matches those priorities. Moreover, Teams should set test credit limits, document harness decisions, and build Monitoring into deployments to catch issues early.

Ultimately, the update makes Copilot Studio more approachable while introducing operational tradeoffs that require planning, and Jensen’s video serves as a practical guide for makers deciding how to adopt the new workflow. Therefore, readers should explore the updated studio, experiment with a controlled test plan, and align governance policies so that benefits like faster authoring do not come at the cost of surprise bills or costly migrations.

Keywords

GitHub Copilot Studio, Copilot Studio tutorial, Copilot Harness explained, GitHub Copilot features, Copilot Studio demo, Copilot Studio setup guide, AI pair programmer Copilot, Copilot Studio best practices