"Don't Build Agents, Build Skills Instead"

Sound familiar... Claude Code is the universal interface, the coding agent has many use cases. Barry Zhang and Mahesh Murag from Anthropic, who helped develop "Agent Skills", argue that the future of AI lies in modular, shareable capabilities rather than monolithic agent development. They outline how organizing procedural knowledge into files allows agents to gain domain expertise and evolve through continuous learning.

I found this Anthropic talk on YouTube, interesting because its core argument closely matches the design logic behind ICM: don’t keep rebuilding specialized agents; make the operational knowledge reusable, inspectable, and easy to improve.

Barry Zhang and Mahesh Murag describe an Agent Skill as an organized folder of procedural knowledge: instructions, references, scripts, and assets, that an agent can load when needed. Their key point is that a generally capable agent often lacks domain-specific expertise; skills package that expertise in a portable form. The agent only loads metadata until the skill is relevant, then pulls in the fuller instructions and supporting files. That keeps context focused while making capabilities composable.

The connection to ICM is strong. ICM uses filesystem structure and Markdown to orchestrate focused, multi-stage work: each stage receives only the context it needs, produces a visible intermediate output, and hands that output to the next stage. Both approaches treat folders and plain-language files as the control surface rather than burying workflow knowledge in opaque code or a giant prompt.

The main difference is scope. Anthropic Skills are reusable capability packages “how to do a kind of work.” ICM is a human-reviewable workflow architecture, “how this particular job moves through stages.” They fit together naturally: an ICM stage could invoke a relevant skill, while the ICM workspace controls sequencing, scoped context, artifacts, and review points.

Put simply:

Skills provide reusable expertise.
ICM provides interpretable orchestration.
MCP provides connections to external tools and data.
Scripts handle deterministic mechanics.

The most important shared idea is compounding learning. Rather than correcting the same behavior every session, capture the improvement in a durable artifact; an instruction, reference, script, or stage contract, so future runs begin better equipped. The talk calls this transferable procedural knowledge; ICM makes it visible, editable, and traceable through the workspace.

One caveat: neither approach replaces every multi-agent framework. ICM is especially suited to sequential, repeatable workflows where humans should inspect stage outputs; highly concurrent or dynamically branching systems may still need fuller orchestration infrastructure.

The talk’s “MCP connects; skills provide expertise” distinction maps especially cleanly to ICM’s stage-specific context and local scripts.

Sources

Don't Build Agents, Build Skills Instead – Barry Zhang & Mahesh Murag, Anthropic

AI Engineer, AIE/Code presented by Google DeepMind.

Anthropic: Barry Zhang, Mahesh Murag, Member of Technical Staff.

Interpretable Context Methodology: Folder Structure as Agentic Architecture (ICM)

Jake Van Clief, David McDermott

Footnotes

A skill and MWP/ICM are related but not equivalent. A skill is a reusable agent capability, typically instructions plus optional scripts, references, and templates, that tells an agent how to perform a class of tasks. MWP (Model Workspace Protocol), described in Van Clief and McDermott’s Interpretable Context Methodology, is a filesystem-based workflow architecture: folders define stages, Markdown carries stage-specific context and contracts, and scripts perform mechanical work. A skill can be a component used within an MWP workspace, but it is not itself the workflow protocol. See the paper on arXiv and its Connected Papers graph.

10 comments

"Don't Build Agents, Build Skills Instead"