Agents That Build Agents: A SKILL-first Blueprint (Microsoft)
A Microsoft (Azure Developer Community) blueprint for “agents that build agents” organized SKILL-first — the agent-skills unit is the primary building block — on the Microsoft Agent Framework SDK and Microsoft Foundry (microsoft‘s agent/model platform). The community-hub post is JS-gated, so this page is rebuilt from its fetchable companions: the same author’s ITNEXT cross-post and, for the authoritative mechanics, Microsoft Learn’s Use skills with Microsoft Foundry agents doc.
The two-layer architecture
A Coding Agent (build-time) sits above a Runtime Agent (the one the business operates), with
the Microsoft Agent Framework as the SDK that ties them together and Microsoft Foundry as the platform
both publish to and run on. The SKILL is the artifact that bridges them: the Coding Agent reads a
SKILL.md before writing code, turning requirements into artifacts that match the team’s framework,
conventions, and fixtures. A skill is versioned with the codebase — update the SKILL once when a
framework idiom changes and every agent built afterward picks it up, which is the pitch for killing
convention drift. This is the self-improving / meta-skill pattern (self-improving-agents) made
concrete in a vendor stack: an agent whose job is authoring the skills other agents run on.
It adopts the open spec — the open question resolves
The standing synthesis question — does Microsoft adopt the open agentskills.io
spec or ship a proprietary format? — now has an authoritative answer. Microsoft Learn states plainly
that Foundry skills “follow the [Agent Skills] specification format” (agentskills.io): a SKILL.md
with name/description YAML front matter, body as injected instructions, one skill per subdirectory.
The runtime uses the spec’s progressive-disclosure pattern via an AgentSkillsProvider — advertise
(inject names+descriptions into the system prompt), load (fetch the full body only when the agent
deems a skill relevant), read resources (pull supplementary assets on demand) — the same load-on-demand
mechanism agentskills-spec defines. Skills are delivered two ways: direct injection (download
SKILL.md into the project; read at startup) or attached to a toolbox, where they surface as MCP
Resources so any MCP client (GitHub Copilot, Claude Code, a custom harness) discovers them via
resources/list/resources/read — following the Skills extension for MCP (SEP-2640). Foundry adds
its own versioned Skills REST API (immutable SkillVersions, a promotable default_version) on top.
So this is convergence, not fragmentation: Microsoft joins anthropic (the spec author) and google (adk) on the same open skill format — and wires it to MCP as the discovery channel, the cleanest evidence yet that skills-as-portable-markdown is an industry attractor rather than one vendor’s house style.
Why it belongs here
Three independent major vendors now ship skill-first agent frameworks on a shared open spec, with MCP as the transport — the strongest single corroboration of this wiki’s standardization thesis. The build-time/run-time split also sharpens the agentic-coding-harness story: the harness is what an agent runs in, the SKILL is what it reads first, and “agents that build agents” is the harness turned on itself.
Related
microsoft · agent-skills · agentskills-spec · anthropic-skills · adk · self-improving-agents · model-context-protocol · agentic-coding-harness