Language: English | 日本語
What this directory is. A practical navigator for routing a piece of work to the architecture that preserves accountability distribution. Together with the ten ADRs, the four quadrants form the two-axis structure of this repository: the ADRs answer per-question (what should be constrained, who is responsible); the quadrants route the work to where those answers apply.
Phase is independent of Quadrant. Each Quadrant can appear in either the design phase or the operation phase of a deployment. ADR-0010 adds a narrow rule about one placement (operation-phase Quadrant 4) without partitioning Quadrants by phase.
The four quadrants were introduced in the 2026-04-29 essay and given
their accountability-attribution treatment in the 2026-04-30 essay.
The naming used in this repository (Algorithmic Search, Autonomous
Agentic Loop) replaces the source essays' provisional names
(Classical AI Quadrant, ReAct Quadrant) with harness-neutral labels;
the source names are preserved as aliases in ../glossary.md.
| Pre-defined workflow | Exploratory | |
|---|---|---|
| Deterministic | (1) Script Quadrant | (2) Algorithmic Search Quadrant |
| Semantic-judgment | (3) LLM Workflow Quadrant | (4) Autonomous Agentic Loop Quadrant |
- Horizontal axis — Can the work be expressed as deterministic rules, or does it require semantic judgment that an LLM provides?
- Vertical axis — Is the control flow of the work pre-defined (the next step is decided by code written in advance), or is it exploratory (the next step must be decided at runtime, in light of what was just observed)?
The vertical axis maps directly to Anthropic's "Building Effective Agents" distinction between predetermined code paths and LLM dynamically directs its own process.
Deterministic × pre-defined. Scripts and pipelines without LLMs. Form filling, data normalization, lookup, validation. AAP's ten ADRs are largely out of scope here — classical software engineering applies.
Deterministic × exploratory. Classical search, dynamic programming, MCTS, reinforcement learning. Routing optimization, scheduling, combinatorial allocation. Out of scope for this repository's LLM-focused ADRs; the concerns are model validation, search-space correctness, and standard SE accountability.
Semantic-judgment × pre-defined. Load-bearing property: the execution path is decided in advance — by humans, by code, by the surrounding workflow — and the LLM is called as a single, bounded step within that path. The LLM does not decide the next action. Each LLM call's contribution is post-hoc separable as a consequence of the load-bearing property; redirect succeeds.
This quadrant divides naturally by input/output modality into two sub-forms, but the load-bearing property is the same in both:
- (3a) Conversational sub-form — specialized chat agents (legal-consultation assistants, diagnostic-support assistants, internal-FAQ systems, expert-knowledge support tools). The human in the conversation is the judging agent; the LLM contributes knowledge retrieval and organization.
- (3b) Batch sub-form — single-purpose LLM functions inside an otherwise deterministic pipeline (invoice matching, ticket triage, exception classification on top of RPA, address normalization). The pipeline owns the control flow; the LLM is called as a fixed role at each semantic-judgment point.
Continuous with the workflow patterns documented in Anthropic's "Building Effective Agents" — prompt chaining, routing, parallelization, orchestrator-workers, evaluator-optimizer — and OpenAI's "A Practical Guide to Building Agents" (manager pattern, decentralized pattern). These map predominantly to the batch sub-form; routing also serves at the top of conversational sub-forms.
The default for most current LLM applications. The 2026-04-30 essay argues that this quadrant has been chronically routed to (4) by elimination because the industry vocabulary lacks a positive name for it.
Semantic-judgment × exploratory. The LLM decides each next step at runtime — the "ReAct" pattern (Yao et al. 2022) and its descendants (Deep Research, exploratory coding agents, open-ended browsing). Legitimate when the work genuinely requires open-ended exploration; the path is not predictable in advance, and bounding the LLM call's role would prevent the work from being done.
Choosing this quadrant commits the deploying organization to a
non-removable attribution gap: when judgment elements blend at
runtime, post-hoc separability is foreclosed, and redirect cannot
succeed at the level of separable contributions. See ADR-0009
Triage Before Autonomy and
../glossary.md#attribution-gap.
Read in order:
decision-tree.md— a five-question triage for routing a piece of work to one of the four quadrants and the ADR set that applies.governance-mapping.md— a dense table mapping each quadrant to its applicable ADRs, required controls, and redirect path.case-studies.md— concrete examples (FAQ classification, research assistant, scheduling, contract review, invoice matching, customer support chat) routed to quadrants and architectures.anti-patterns.md— common ways the quadrant choice goes wrong, why each fails, and how to recover.
The two artifacts answer different questions:
- The ten ADRs answer "what should be constrained, and who is responsible" given that we are operating in a quadrant where accountability distribution is meaningful.
- The four quadrants answer "is accountability distribution meaningful for this work, and if so, in which architectural regime?"
The two-axis posture preserves AAP's "ten judgments, not a fixed framework" stance: the quadrants are a diagnostic frame for routing, not a prescriptive taxonomy. Different quadrants apply different subsets of the ADR set with different load-bearing weight; the ADR set itself is not partitioned by quadrant.
Phase (design vs operation) is independent of Quadrant. Every Quadrant can appear in either phase: a workflow prototyped in design and promoted to operation traverses both phases of the LLM Workflow Quadrant; an autonomous loop used to explore a domain (design) before its findings are encoded into a workflow (operation) traverses both phases of Quadrant 4 → Quadrant 3.
ADR-0010 surfaces the Phase-crossing decision for one specific placement: operation-phase Autonomous Agentic Loop Quadrant deployments. There, the deploying team must record whether new patterns surfacing in operation are handled dynamically in place or routed back to design as feedback. The rule is procedural, not architectural — it does not partition Quadrants by Phase.
This layer is part of AAP's active publication and will be updated as adoption experience accumulates. The vocabulary (especially the boundary between LLM Workflow Quadrant and Autonomous Agentic Loop Quadrant for multi-turn chat agents) is still being refined; see ADR-0009's open questions.