Anthropic masterclass: agent harnesses

Anthropic used a new masterclass video to put the focus on the machinery around coding agents rather than on raw model performance. The session, published on YouTube on May 21, 2026, centers on “agent harnesses” for large codebases — the routines, constraints and checks that let a model work through a repository without drifting or breaking things. The framing is narrower than a general product demo and more operational than a benchmark discussion. The video’s premise is that coding agents in large repositories need more than access to a strong model. The material emphasizes how an agent explores a repo, how much code and documentation it pulls into context, how it validates changes and where a person steps in. In that account, the harness is the working system that surrounds the model. ### Why did Anthropic focus on the harness instead of the model? Anthropic’s session describes coding work in large codebases as a sequence of bounded steps rather than a single generation task. The video highlights planning loops, file selection, test execution and recovery behavior as parts of the system that determine whether an agent can complete a task reliably. That framing shifts attention from model IQ to operating method. In the masterclass, capability is presented as something produced by the combination of model, tools, constraints and evaluation, not by the model alone. ### What does “repo exploration” mean in practice? Large repositories contain far more code than a model can inspect at once, and the session treats navigation as a first-order problem. The material points to strategies for finding the right files, reading documentation, tracing dependencies and narrowing the search space before making edits. Context budgeting sits alongside that process. The video emphasizes choosing what to show the model and what to leave out, so the agent does not waste context on irrelevant files or lose the thread of the task in a large codebase. ### Where do testing and recovery enter the picture? Testing appears in the session as a gate on agent action, not as an afterthought. The material stresses validating changes, checking whether edits actually solve the stated problem and using test feedback to decide the next step. Recovery logic is treated as another core part of the harness. If an edit fails, if a test breaks, or if the agent heads down the wrong path, the system needs a way to retry, back up or hand the task off for review rather than continue blindly. ### Why are human review points part of the design? Human review shows up in the video as a control point inside the workflow. The session describes moments where a person may need to approve a plan, inspect a risky change or decide whether the agent should continue after an uncertain result. That approach keeps the agent inside a supervised loop. In Anthropic’s presentation, the review step is part of the harness design for code that matters, especially when the repository is large and the consequences of a bad change are harder to contain. ### What is the broader takeaway for coding agents? The masterclass presents coding agents as systems that need orchestration around them. The key ingredients in the session are planning loops, execution limits, repo navigation, test-based validation and explicit handoffs between model and human. The video remains available on YouTube at Anthropic’s published link, where the full session lays out the harness concepts in more detail and alongside the company’s other coding-agent material.