A methodology is agent-extensible only where it is closed under its own recommendations
Type: kb/types/note.md · Status: seedling · Tags: foundations, constraining
An LLM can consume a theory and act on it, so encoding a good methodology is one way of building capability: point an agent at the methodology and it executes. The stronger move is that the agent can also extend the system the methodology governs — writing the code, schema, or validator the methodology calls for. But that self-extension is not free-floating. An agent can extend a methodology autonomously only where the methodology specifies the decisions its own extension requires. Call this closure under its own recommendations.
The condition
Extending a methodology-governed system forces two meta-decisions:
- Representational form — should the next artifact stay prose to be interpreted, or be frozen into deterministic code, schema, or grammar? A methodology is closed on this axis when it hands the agent the criteria to decide, rather than leaving the agent to guess. Commonplace supplies these as the codify-versus-LLM decision heuristics and the constraining gradient from convention to code; the decision itself is codification.
- Verification — once the artifact exists, how does the agent know it is correct? A methodology is closed on this axis when it tells the agent what oracle to build or invoke.
Where both are specified, the loop closes: the agent reads the methodology, recognizes a case, produces the artifact the methodology prescribes, and verifies it — all without stepping outside the methodology. Where either is missing, the agent must improvise the meta-decision, and improvised meta-decisions are exactly where two sessions diverge — because natural-language instructions are interpreted, not executed.
Verification is the ceiling, not understanding
Closure makes self-extension possible; it does not make it unbounded. The binding constraint is that the agent can produce an artifact only as reliably as it can check that artifact, since the boundary of automation is the boundary of verification. A methodology closed on representational form but open on verification will generate artifacts the agent cannot confirm — output, not automation. This reframes what a methodology's verification machinery (typed artifacts, validators, review gates) is for: it is not bureaucracy around the knowledge, it is what raises the ceiling on how far an agent can extend the system from the methodology alone.
So the power of an agent-executable methodology is not set by how much the agent understands — understanding is assumed — but by two properties of the methodology itself: how completely it specifies its own extension decisions, and how cheaply the results of those decisions can be verified.
Why the artifact is retained, not re-derived
Closure explains how an agent produces a codified artifact from the methodology; it does not argue the agent should reproduce it each session. A persisted symbolic artifact is deterministic and inspectable in a way a re-derivation is not — re-deriving pays the cost again and risks two interpretations drifting apart. This is the reason an agentic system keeps a knowledge base of committed artifacts rather than a single theory document the agent re-expands on demand: the methodology generates the artifact via closure, and the KB retains it so nobody re-derives it. In agent systems the prescription/implementation boundary collapses — the prescription and the code it becomes are the same retained thing at different points on the constraining gradient.
Scope
- Closure is a direction, not a binary. No real methodology fully specifies every extension decision it could face; the claim is that agent-autonomy scales with how much it specifies, and stalls at the first meta-decision it leaves open.
- The counter worth taking seriously. A capable agent brings general competence and can improvise the reason-vs-codify and verification decisions the methodology omits. Where that improvisation is reliable, closure buys less. The claim's force therefore tracks how consequential and divergence-prone the omitted meta-decisions are — high for what-to-codify and how-to-verify, low for cosmetic choices.
- This is a property of the methodology-as-agent-input, not of any one system. Commonplace is the worked example, not the claim; the principle should hold for any methodology handed to an agent to execute and extend.