Agency
A self-hosted engine for composing AI agents from reusable natural-language building blocks called primitives — role components, desired outcomes, and trade-off configurations. When an agent underperforms, you change one primitive instead of rewriting the entire prompt. Runs as a background MCP service. Built by Vaughn Tan (@arbois).
Key Quotes
"A monolithic agent has no user-serviceable component parts to individually fix or replace."
The thesis statement. This is the same argument that drove the Unix philosophy, microservices, and modular programming — applied to the prompt layer. Agency treats prompts as software components: decompose, test, recompose.
"When something does not work, you change one primitive. Everything else stays the same."
The surgical-improvement guarantee. Every primitive addition potentially improves all future agents, not just the current task. This is the compounding loop that makes the overhead of composition worthwhile.
"Memory about agents, not memory within them."
The three knowledge layers — primitives, compositions, performance — persist at the system level. No individual agent carries context forward. This is a clean separation of concerns: Agency owns identity and composition; the runtime owns execution and persistence.
The Case Study That Matters
During v1.2.3 development, an initial PRD review agent scored 26/60 — it read the spec in isolation but couldn't check against the actual codebase. After adding a single one-sentence primitive for codebase-awareness (checking API signatures, SQL validity, testability), a re-run scored 55/60. The composed agent caught a field name mismatch (similarity vs. score) that would have silently broken every relevance computation.
Across four review rounds, composed agents caught 14 implementation blockers that general-purpose Claude agents missed.
This is the strongest evidence in the repo. It's not a benchmark — it's a real development workflow where the modular approach demonstrably outperformed monolithic prompting.
Key Themes
Critical Analysis
What's genuinely novel: The three-type taxonomy (roles, outcomes, trade-offs) is more specific than "break prompts into pieces." Trade-off configurations in particular acknowledge that agent design involves value conflicts — thoroughness vs. speed, exploration vs. focus — and that these should be explicit, not buried in prose. The evaluation loop closes: compose → execute → score → improve primitives → recompose. This is Compound Engineering applied to prompts themselves.
The license is a adoption ceiling. Elastic License 2.0 prohibits commercial hosting, which blocks the most natural deployment pattern (a shared team service). This may be intentional — Vaughn Tan appears to be building a research-informed tool, not a SaaS business — but it caps the potential user base.
Scope honesty is a strength, not a weakness. Agency explicitly disclaims long-running autonomous agents. It composes, hands off, and gets out of the way. This makes it complementary to runtimes like Slate (thread-and-episode architecture), Cord (dynamic task trees), and Serf (non-interactive coding agent). Agency handles "what agent should do this?" — the runtime handles "keep doing it."
The primitive extractor is the unspoken differentiator. The README mentions a bundled tool that pulls primitives from existing workflows, skills, and sessions. If this works well, it solves the cold-start problem that kills most component libraries. Without good primitives, the composition engine is just overhead.
The tension with minimalism: What I learned building an opinionated and minimal coding agent argues that four tools and no MCP is sufficient because frontier models already understand coding agents through RL training. Agency bets the opposite direction — that explicit, curated primitives outperform implicit model knowledge. The 26→60 case study is a data point for Agency's side, but it's one case study on one codebase.
The real question is whether primitives are genuinely reusable across domains. Role components like "identifies logical gaps" might transfer. "Checks SQL schema validity" probably doesn't. The value of the primitive library depends on how much domain-specific vs. domain-agnostic capability exists in agent work. If most useful primitives are domain-specific, the library fragments. If most are domain-agnostic, it compounds.
Sources: Agency — agentbureau/agency Last updated: 2026-05-15