3.5 KiB
Evolution of Agent Tooling (Tools → MCP → Skills)
#concept
Summary
Konstantin's three-generation map of how agents get capabilities: Tools (2022–2024) → MCP (late 2024–2025) → Skills (2025→). Each generation trades off developer cost, context consumption, and flexibility.
Current Understanding
| Generation | What it is | Strength | Weakness |
|---|---|---|---|
| Tools | Function tagged and injected into the model call (LangChain/LlamaIndex) | Direct control | Every tool eats context; you write/adapt each yourself |
| MCP | Locally-deployed server exposing many tools via a standard protocol (Anthropic); marketplaces exist | Ready-made tools, plug-in | Dozens of tools per server → context blowout (>100 = chaos); network-load = security surface; needs a developer |
| Skills | Directory + SKILL.md (+ tools + data), two-stage loading |
Hundreds without context blowout; author with text in any language; share via git/messenger | No standards yet for data/limits |
When to use which: Skills when tasks are unknown/diverse or tool count is ~5–50; MCP when the agent is narrow, tasks are uniform, and the same small toolset applies every time. The line blurs — Claude Code converts MCP servers into skills (file laid down, functions not all injected), erasing most MCP downsides. Konstantin doesn't hate MCP; its problems are largely solved.
A fourth position: skip the progression. thorsten-ball uses none of the three generations as user-authored artifacts — no skills, no MCP servers, no slash commands — and locates capability instead in the codebase plus AGENTS.md plus a rich prompt (2026-07-28-agentic-engineering-10x-developer). amp does ship structure (Oracle/Painter/Puck sub-agents, a model/effort dial), but the vendor curates it, not the user. Read against this table, his claim is that the progression's real axis was never tools → MCP → skills but who supplies the context and where it lives — and that for someone working in one well-tended repo, the repo wins. See the three competing readings logged on skills-as-memory.
Evidence
- Three generations, per-generation problems, skills-vs-MCP decision table, Claude-Code-turns-MCP-into-skills caveat — 2026-07-14-skills-based-on-git.
- Skills as portable markdown folders across Claude/Perplexity/Gemini — 2026-07-14-gap-between-ai-users-irreversible.
- The skip-it-all position; vendor-curated sub-agents in place of user-authored skills — 2026-07-28-agentic-engineering-10x-developer.
Related Pages
- Concepts: skills-as-memory, harness, context-as-scarce-resource
- Tools: claude-code, hermes, amp
- Entity: konstantin, thorsten-ball
Contradictions / Uncertainty
- "Everything changes every 6 months" — MCP was just ratified and A2A is already wanted; this map may shift quickly. Status: tentative.
- The map assumes each generation supersedes the last. Thorsten's practice suggests a parallel track that never enters the table at all (repo +
AGENTS.md), which would make the progression a history of one branch rather than of agent capability as such. Status: tentative.
Next Questions
- Where does agent-to-agent (A2A) sit in this progression? (Partial datapoint 2026-07-28: amp shipped agent-to-agent messaging and a meta-agent that spawns and controls other agents — see async-by-default. It arrived as a harness feature, not as a fourth generation of capability-delivery, which the table would not have predicted.)