Skip to content

Research: skills as decision codifiers — trade LLM encoding for deterministic decision logic #79

@AccessiT3ch

Description

@AccessiT3ch

Research Question

How much of the guidance currently encoded as natural-language instructions in agent files can be converted into deterministic, non-LLM decision logic — reducing the cognitive load on the LLM and increasing behavioral reliability? Can delegation itself be modelled as a reusable skill?

Context

The current agent architecture relies heavily on LLM instruction-following: an agent reads a long body of prose instructions and the model must infer the correct action. This is the expensive, drift-prone path. Every token the LLM spends deciding "should I delegate this?" or "which phase gate applies here?" is a token not spent on the actual task.

The insight: skills can encode decision logic, not just procedures. A skill file that includes a full decision tree ("if task type is X → delegate to agent Y; if output file exists → skip phase; if review returns CHANGES_REQUESTED → fix and re-submit") removes that reasoning from the LLM's cognitive budget at runtime.

Three specific hypotheses to test:

  1. Delegation as a skill: The Delegation Decision Gate in the Executive Orchestrator is currently a prose lookup table. Could this be a SKILL.md that multiple agents load — pinning the routing logic deterministically and removing it from each agent's instruction body?

  2. Phase gate logic as a skill: The per-phase sequence (prune → checkpoint → commit → review → compact) is currently re-stated in each orchestrator instruction set. A phase-gate skill would encode this as a procedure multiple agents share.

  3. Decision tree offloading: What is the theoretical minimum instruction-body size for an agent if all decision logic is extracted into referenced skills? What remains — identity/persona/values — is the irreducible core.

Scope

  • Survey: PDDL and classical AI planning (what can be fully formalized); BDD/Gherkin (precondition/action/postcondition as decision encoding); finite state machines as agent workflow models
  • Audit current agent files for recurring decision patterns (delegation routing, phase gates, review cycles) that appear in ≥2 agent bodies — candidates for skill extraction
  • Design prototype skill(s) for: delegation-routing, phase-gate-sequence
  • Evaluate: what is the instruction-body size reduction from extracting these into skills? Token budget saved?

Gate Deliverables

  • D1 — Audit: list of decision patterns appearing in ≥2 agent bodies; estimated token savings from extraction
  • D2 — delegation-routing SKILL.md prototype with full decision tree for Executive Orchestrator routing
  • D3 — phase-gate-sequence SKILL.md prototype encoding the per-phase sequence as a reusable procedure
  • D4 — Theoretical minimum instruction-body analysis: what is the irreducible core of an agent's instructions?
  • D5 — Findings documented in docs/research/skills-as-decision-logic.md

Related Issues

Metadata

Metadata

Assignees

No one assigned

    Labels

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions