ADPS Agent Systems · Engineering Concepts

Engineering Concept Catalogue

Grouped by engineering concern, with the source of each concept recorded.

These pages describe recurring engineering structures. Start with the key concepts or open a group that matches the current problem. Each page distinguishes the initial practice or workshop proposal from later ADPS naming and formalization.

Key concepts

Control and Narrative Planes

Separate model-readable context from bounded values that drive the runtime.

Execution vs Content Agents

Choose runtime controls from the deliverable, state dependency, and cost of failure.

Mechanical State Plane

Bind exact business values to auditable source coordinates.

Abstraction-Reconstruction Loop

Extract common structure, then restore critical differences to test implementability.

Probabilistic Core, Deterministic Shell

Models handle open judgement; deterministic mechanisms protect production boundaries.

Context Contract

Specify what information may enter a task context, with source, freshness, and budget.

Plan Compilation

Compile a reasoned plan into an executable structure with dependencies, types, authority, and acceptance.

Three Powers of Agency

Inspect agent power at the points of information, judgement, and action.

Browse by problem

System boundaries and state7

Locate control, factual state, and runtime boundaries.

Control and Narrative Planes

Separate model-readable context from bounded values that drive the runtime.

Execution vs Content Agents

Choose runtime controls from the deliverable, state dependency, and cost of failure.

Three Session-State Planes

Coordinate narrative, mechanical business values, and task scheduling without mixing write ownership.

Mechanical State Plane

Bind exact business values to auditable source coordinates.

Orchestrator and Facade

Keep control flow in the orchestrator and input/output presentation in the facade.

Disk Fact Plane

Keep stage facts, progress, and acceptance evidence in recoverable external contracts.

Probabilistic Core, Deterministic Shell

Models handle open judgement; deterministic mechanisms protect production boundaries.

Context and memory6

Control how information enters context, persists across steps, and is retrieved.

Anchor, Ledger, Collection

Preserve the goal, append task facts, and project only the material needed now.

Scratchpad Board

Store the current goal, step, observation, blocker, evidence, and next action in a small working surface.

Memory Envelope

Assemble a versioned context package for one model call.

Layered Memory L1/L2/L3

Separate current input, task facts, and reviewed cross-task experience.

Knowledge Compilation

Transform knowledge sources into traceable, incrementally updated runtime assets.

Context Contract

Specify what information may enter a task context, with source, freshness, and budget.

Reasoning and action mechanisms3

Compare how a model advances work through reasoning, observations, code, and tool calls.

ReAct Reasoning-Action Loop

Alternate reasoning and action so each next step can use a fresh observation from the environment.

Programmatic Tool Calling

Let the model write a bounded program that calls registered tools and reduces intermediate results before they enter context.

CodeAct: Code as Action

Use executable code as the agent's action language for computation, libraries, and tool composition.

Planning, execution, and write-back9

Turn intent into schedulable steps and govern write-back of intermediate and business results.

Intent as Compilation

Translate open language into a validated, bounded control signal with an explicit fallback.

Task DAG and State Machine

Make dependencies, node status, acceptance, and recovery explicit.

Reasoning Assetization

Move repeated, reviewable decisions into versioned capability assets before runtime.

Typed IR Chain

Narrow natural language through typed, verifiable structures before write-back.

One-Diagram Unit of Work

Commit one diagram, one resolved scope, one write set, and one receipt at a time.

Structural Write-Back Gate

Check scope, types, references, completeness, and real change before side effects.

Resumable Clarification Job

Pause the same durable job when one required field is missing, then resume it.

Vocabulary Equivalence Layer

Align planning terms, domain types, and the host metamodel through one tested mapping.

Plan Compilation

Compile a reasoned plan into an executable structure with dependencies, types, authority, and acceptance.

Collaboration and runtime control7

Handle handoffs, concurrency, responsibility, and runtime composition across participants.

Topology Lowering

Compile high-level collaboration semantics into runtime primitives such as serial, parallel, and routing edges.

Topology Governance Matrix

Review serial, parallel, and routed structures across identity, authority, safeguards, and provenance.

Handoff Contract

Transfer goal, artifacts, decisions, authority, responsibility, and acceptance together.

Three Collaboration Planes

Review Agent-Agent, Human-Agent, and Human-Human collaboration around the system.

Hook Composition

Compose orchestration, governance, observation, and recovery at lifecycle events.

Unified Convergence Node

Resolve several reasoning branches into one actionable result under shared evidence and adjudication rules.

Write-Conflict Domain

Identify files, identifiers, configuration, domain objects, and environments that concurrent tasks may jointly mutate.

Evaluation, reflection, and governance12

Define who judges results, how feedback returns, and when the system may continue acting.

Describe how patterns are abstracted from practice and tested again in concrete systems.

Abstraction-Reconstruction Loop

Extract common structure, then restore critical differences to test implementability.

‘Initial source within ADPS’ identifies the earliest record in ADPS material; it is not a claim of global invention.