Inscription Architecture

A formal specification for how agents receive instructions, match patterns, create new patterns, propagate policies, and communicate. This extends the current HAAK agent architecture (fixed roles,…

A formal specification for how agents receive instructions, match patterns, create new patterns, propagate policies, and communicate. This extends the current HAAK agent architecture (fixed roles, manually scaffolded) toward a self-modifying system where agents learn and share behavioral patterns.

#Relationship to existing architecture

The current HAAK system uses manually authored patterns (skills, processes, policies). This spec describes how agents could create, search for, and evaluate patterns autonomously — the "learning etc" phase from the [[Assessment of Priorities]] roadmap.

Mapping to current types:

  • Pattern = generalized name for process / policy / protocol / skill
  • Constitution = root policy (already implemented in process-architecture)
  • Policy = cross-cutting constraint (already implemented)
  • Instructor/instructee = new role model extending current fixed agent types

#A. Pattern matching and execution

An agent receives an instruction (prompt):

  1. Check if any existing pattern matches the instruction
  • Match could be specified in many ways
  • Analogous to Claude/OpenAI skill matching
  1. If match found → use the pattern to complete the instruction
  2. If no match → ask for feedback:

a. Stop the task (someone else does it) b. Request clarification → check again (1) c. Create a new pattern (B)

When an agent needs a new pattern:

  1. Search for analogues:

a. Own library of previously used patterns b. Other agents' pattern libraries (see C) c. External patterns (web search) d. Homology / metaphor from other domains e. Instructor's guidelines/policy (B.1.f)

  1. Reason about the needed pattern based on policy + analogy
  2. Propose a new pattern to complete the task

#B.1.f: Instructor policy propagation

The instructor (requesting agent) passes guidelines that:

  • Are implicit in the instruction
  • Are rules/guidelines of the most general sort
  • Apply to every new pattern the instructee creates
  • Need not be repeated every instruction if the agent remembers instructor policy
  • Can be passed like a system prompt (limits to per-use scope otherwise)

Terminology:

  • Instructor = agent making the request
  • Instructee = agent creating/learning the new pattern

#C. Policy passing between agents

An agent requesting help passes its policies to the receiving agent (like a system prompt — rules, not specific instructions).

An agent may refuse or request clarification if:

  • It checks patterns and finds no match (→ A.3)
  • Requester may then: try another agent type, change the request, clarify, or escalate

Role dynamics: An agent may act as requester, instructor, executor, or instructee at different times. These are roles, not types — unless precluded by policy.

#D. Policy framework

#D.1–D.4: Policy hierarchy

  1. Constitution: rules followed by all agents (requires human approval to change)
  2. Agent-specific policies: rules specific to an agent's capabilities; any pattern created/used must adhere
  3. Policies may be shared across agents; they define/limit domains of action
  4. During cooperation: policies of both requester AND executor must be followed

#D.5: Policy creation and amendment

a. Constitution changes require user/human approval b. Agent-created policies should be human-checked c. A policy is generally a generalization of a successful pattern or set of patterns d. Proposed policies must be checked for:

  • Adherence to constitution
  • Non-collision with other policies
  • Non-contradiction

e. Policies should be atomic (~1 sentence + explanation) for composability f. Automatic periodic policy review for non-contradiction, especially with constitution

#D.6: Communication between agents

a. Communication by request or self-initiation (if policy/instruction allows) b. All communication recorded — external, in both agents' output, auditable c. Communication may be on-the-record (in agent memory/patterns) or off-the-record (current session only; transcripts not part of agent memory by default) d. Group communication (>2 agents): entire transcript saved to each e. Limit communication to preserve:

  • Independent views
  • Prevent leakage / conformity
  • Support adversarial processes

f. Agents may review other agents' performance based on task output (part of pattern evaluation)

#Connection to inscription project

The Library Theorem (inscription) formalizes why external memory (patterns, policies) enables capabilities beyond in-context reasoning. This spec is the operational counterpart: how agents actually use the external library. A learning-based inscription agent would implement sections A-B (pattern matching + creation) with the formal guarantees the theorem provides.

#Open questions

  • Is a shared Turing tape/machine with multiple read/write heads equivalent to or super-Turing? [Lit search needed]
  • How does pattern evaluation work? (Referenced as B.5 but not yet specified)
  • When does a pattern graduate to a policy? (Relates to process-architecture maturity model)
  • How is "match" specified? (Embedding similarity? Keyword? Structural?)

haak · created 2026-02-22 · source: handwritten notes Feb 17-19 2026 · zach

Architecture 05 — Inscription Architecture — 2026 — Zachary F. Mainen / HAAK