Language: English | 日本語

Case Studies — Quadrant Routing in Practice

Purpose. Concrete examples of routing common business AI work to a quadrant — with the architecture, applicable ADRs, and (where relevant) typical phase placement spelled out. The cases below are illustrative, designed to stimulate the team's thinking, not to be pattern-matched. Phase ≠ Quadrant: the same Quadrant can appear in both design and operation phases, and the cases note which placement is typical, not exclusive.

Currency. Each case carries a publication date. The choices reflect what is sensible given current LLM and tooling capability as of that date. Re-evaluate when capabilities shift.

(Autonomous Agentic Loop Quadrant, Typical placement: Design phase) Coding agent

Published 2026-05-01.

• Work. A developer asks a coding agent to investigate, modify, test, and commit code in a repository. Each session ends when the task is complete.
• Phase. Design phase resident. Even when the agent is "always on" (Devin, GitHub Copilot Coding Agent in CI/CD), each session is an independent design session that completes when its exploration finishes; state does not accumulate across sessions in operation-relevant ways.
• Quadrant. Autonomous Agentic Loop Quadrant.
  • Q0 (phase?) — Design.
  • Q1 (deterministic?) — No.
  • Q3 (workflow definable?) — No; the next action depends on what the previous file read or test result returned.
  • Q4 (gap-bearer + Phase-crossing) — Both must be answered. The gap-bearer is typically the developer who initiated the session; for unattended CI runs, the gap-bearer is the repository owner. The Phase-crossing decision is automatic (this is the design phase).
• Architecture. ReAct-style autonomous loop with a controlled tool set (file read, file edit, shell, test runner) and an explicit termination criterion (task complete / approval gate).
• Applicable ADRs. All ten. ADRs 0005, 0006, 0008, 0009 carry weight; ADR-0010's Phase-crossing condition is satisfied automatically.
• Why this is not an exception to ADR-0010. Coding sessions are sequences of independent design sessions. Their continuous availability is a deployment-level fact, not a phase-level fact. Treating coding agents as operation-phase deployments is the category error ADR-0010 names.

(Autonomous Agentic Loop Quadrant, Typical placement: Design phase) Deep Research for ad-hoc analysis

Published 2026-05-01.

• Work. A user submits a research question whose answer requires iterative search, retrieval, and synthesis across an unbounded knowledge corpus. Each query ends when the synthesis is complete.
• Phase. Design phase resident (per-query design session).
• Quadrant. Autonomous Agentic Loop Quadrant.
• Architecture. ReAct-style autonomous loop with retrieval tools (search, fetch, cite) and a synthesis step.
• Applicable ADRs. All ten. The session-level gap-bearer is the user (or the user's employer for paid deployments); the Phase-crossing decision is automatic.
• Why this is not operation-phase. A query session is exploratory; the next search depends on what the previous result returned. The deployment's appearance of being "operation phase" (the service is always available) is a delivery property, not a phase property.

(LLM Workflow Quadrant batch sub-form, Typical placement: Operation phase) FAQ classification

Published 2026-04-30.

• Sub-form. Batch sub-form: the pipeline owns the control flow; the LLM is invoked once per question with a fixed role (classify(question) -> category).
• Work. A user submits a question; the system classifies it into one of N FAQ categories and returns the matching answer.
• Phase. Operation phase. Default composition (no Phase-crossing decision needed; Quadrant 4 is not in the composition).
• Quadrant. LLM Workflow Quadrant.
  • Q1 (deterministic?) — No, semantic judgment required.
  • Q3 (workflow definable?) — Yes; one classification call, one lookup, one return.
• Architecture. Single-purpose LLM function: classify(question) -> category with a documented schema. Followed by a deterministic lookup category -> answer. No autonomous loop.
• Applicable ADRs. 0001, 0003, 0006, 0007 (full list in governance-mapping.md).
• Redirect path. If the classification is wrong, the role owner of classify() (model selection, prompt design) is the redirect target.
• Anti-pattern. Implementing as an autonomous agent that "decides which FAQ to retrieve." See anti-patterns.md#bounded-work-on-autonomous-loop.

(LLM Workflow Quadrant conversational sub-form, Typical placement: Operation phase) Customer support specialist chat

Published 2026-04-30.

• Sub-form. Conversational sub-form: the human in the conversation is the judging agent; each turn is a bounded LLM call (system prompt + retrieval + history) whose role is fixed by the surrounding harness.
• Work. A multi-turn chat with a user, grounded in a product knowledge base. Escalates to a human when the conversation reaches a topic outside the scope.
• Phase. Operation phase. Default composition.
• Quadrant. LLM Workflow Quadrant (multi-turn but bounded — see the boundary discussion below).
  • Q1 (deterministic?) — No.
  • Q3 (workflow definable?) — Yes: each turn is a bounded LLM call (system prompt + retrieval + history) followed by an escalation check.
• Architecture. Specialist chat agent: a single LLM call per turn with a fixed role (answer questions about product X using retrieved context). The escalation check is itself a bounded classification call.
• Applicable ADRs. 0001, 0003 (RAG content is untrusted), 0006, 0007.
• Redirect path. Wrong answer → role owner of the answer call (RAG curator if the wrong document was retrieved; prompt designer if the prompt encouraged the wrong synthesis).
• Boundary note. Multi-turn chat agents sit on the boundary between (3) and (4) — see ADR-0009 open question 1. The test: does the role of each turn's LLM call stay bounded under unusual user input, or does the agent start "deciding what to do next" beyond answering questions about product X? If the role bound holds, this is (3). If not, it has drifted into (4) and the ADR-0009 triage applies.

(LLM Workflow Quadrant batch sub-form, Typical placement: Operation phase) Invoice matching

Published 2026-04-30. Source: 2026-04-29 essay.

• Sub-form. Batch sub-form: the pipeline (process_invoice()) owns the control flow; a single-purpose LLM function (match_line_items(...) -> Verdict) is invoked at the one semantic-judgment point.
• Work. Match incoming invoices against purchase orders; approve, escalate, or reject.
• Phase. Operation phase. Default composition.
• Quadrant. LLM Workflow Quadrant.
  • Q1 (deterministic?) — Mostly yes (PO existence, expiry, duplicate, exact-match amount), but the line-item matching step requires semantic judgment.
  • Q3 (workflow definable?) — Yes; control flow is a deterministic pipeline with one LLM call at the line-item-matching point.
• Architecture. Deterministic pipeline (process_invoice() in the source essay) with a single-purpose LLM function match_line_items(invoice_lines, po_lines) -> Verdict.
• Applicable ADRs. 0001, 0003, 0005 (changes to the matching prompt/schema go through approval), 0006, 0007.
• Redirect path. A wrong match isolates to either the LLM function's role owner or the pipeline's logic owner. Both are identifiable.

(Autonomous Agentic Loop Quadrant, Typical placement: Design phase) Research assistant for open-ended literature exploration

Published 2026-04-30; phase note added 2026-05-01.

• Work. A user submits a research question whose answer requires iteratively searching, reading, and synthesizing across a knowledge corpus whose structure cannot be enumerated in advance.
• Phase. Design phase resident (per-query design session, like the Deep Research case above; included separately because it documents the team-deployed variant rather than the consumer-tool variant).
• Quadrant. Autonomous Agentic Loop Quadrant.
  • Q0 (phase?) — Design (per query).
  • Q1 (deterministic?) — No.
  • Q3 (workflow definable?) — No; the next search depends on what the previous search returned. Bounding the call's role would prevent the work.
• Architecture. ReAct-style autonomous loop (Thought → Action → Observation), with a controlled tool set (search, retrieve, cite) and an explicit termination criterion.
• Applicable ADRs. All ten. ADRs 0005, 0006, 0008, 0009 load-bearing; ADR-0010's Phase-crossing decision is satisfied automatically (design phase).
• Required commitments at deployment time.
  • ADR-0009. The triage decision is recorded; the work is explicitly named as Autonomous Agentic Loop Quadrant work.
  • ADR-0008. The accountable operator is named. Alice is the accountable operator for this research assistant deployment.
  • Pre-named gap-bearer. Same operator (or a higher-level role if the work crosses individual authority).
  • ADR-0005. Behavior-modifying writes (changes to the loop's prompts, tools, or termination criteria) pass through Alice's approval gate. As a deployment-specific choice — stricter than the ADR floor — Alice also gates outward-facing actions (publishing the report, sending email, posting to a channel).
• Redirect path. Component-level redirect is foreclosed. If the report is wrong, Alice is the redirect target (per the gap-bearer commitment). Post-incident options: retrain, re-architect (move parts of the work to LLM Workflow Quadrant), retire.

(Algorithmic Search Quadrant) Resource scheduling

Published 2026-04-30.

• Work. Allocate a finite set of resources (rooms, machines, staff) across a finite set of demands subject to constraints.
• Quadrant. Algorithmic Search Quadrant.
  • Q1 (deterministic?) — Yes; the constraints and objective are fully specifiable.
  • Q2 (search space static and finite?) — No; combinatorial enumeration is required.
• Architecture. Constraint solver, integer programming, or metaheuristic search. No LLM in the loop.
• Out of scope for AAP. Standard SE accountability applies; the algorithm author and the constraint specification owner are the redirect targets.
• Anti-pattern. Implementing as an autonomous agent that "asks an LLM to schedule." See anti-patterns.md#or-problem-on-autonomous-loop.

(Hybrid) Contract review

Published 2026-04-30.

• Work. Review a contract: extract clauses, check against a policy library, flag novel-pattern clauses for human attention.
• Quadrant. Hybrid.
  • Clause extraction — LLM Workflow Quadrant. Each clause type is a bounded call with a documented schema.
  • Policy comparison — LLM Workflow Quadrant. Each policy check is a bounded call.
  • Novel-clause analysis — Autonomous Agentic Loop Quadrant only if the analysis genuinely requires open-ended exploration of legal precedent. In practice, this part is often handed to a human rather than an autonomous loop, because the cost of a hallucinated precedent citation is high.
• Architecture. Pipeline of LLM Workflow Quadrant calls, with an approval-gated escalation to either a human or (with ADR-0008/0009 satisfied) an Autonomous Agentic Loop for the novel-clause portion.
• Applicable ADRs for the LLM Workflow portion. 0001, 0003, 0005, 0006, 0007.
• Applicable ADRs for the Autonomous Agentic Loop portion (if used). All ten.
• Boundary discipline. The handoff between the two portions is itself a structural choice and is logged. The Autonomous Agentic Loop portion has its own named gap-bearer, distinct from the workflow operator if the role separation matters.

(Autonomous Agentic Loop Quadrant, with the operator-experience caveat) Knowledge-base copilot

Published 2026-04-30. Source: implementation experience cited in the 2026-04-29 essay.

• Work. A vast official documentation set; users ask questions whose answers require traversing the knowledge graph in ways that cannot be planned in advance.
• Quadrant. Autonomous Agentic Loop Quadrant.
  • Q3 (workflow definable?) — No. The next search depends on observations. Bounding would prevent the work.
• Architecture. ReAct-style loop with retrieval tools.
• Applicable ADRs. All ten.
• Reported operator experience. Strong results on the work itself; reduced predictability under operation. The number of iterations and the path the loop takes vary across runs. This is the intrinsic property of the quadrant, not a tooling limitation. Plan for it.
• Required commitments. ADR-0008 named operator + ADR-0009 named gap-bearer. ADR-0005 approval gate on behavior-modifying writes; gating external publication of each constructed answer is a deployment choice beyond the ADR floor (some deployments treat each answer as itself an approval-gated artifact; others treat the deployment envelope as approval-gated and let answers stream).

(Skill-design gradient case) Same job, different phase, different shape

Published 2026-05-02. Source: 2026-05-02 essay (essay 7).

• Work. Distill recurring patterns from a body of natural-language or structured-text artifacts into reusable, fixed units (rules, templates, embeddings, classification schemes). The same job recurs at different points in the lifecycle — once during exploratory authoring, once after the artifact set has settled.
• The same job, two phases, two shapes. The job's position on the skill-design gradient changes with phase, not with model capability. Two implementations of the same job coexist legitimately:
  • Design phase form. A natural-language skill that an LLM invokes at runtime. The target codebase, the IDE, the relevant file types, and the execution environment vary across invocations; fixed paths and naming conventions cannot be assumed in advance. Sits closer to the Autonomous Agentic Loop Quadrant in shape (per-invocation runtime judgment), but is still a single skill, not a free-form ReAct loop. Target identifiability is low; scale-resilience is not the constraint because scale is small.
  • Operation phase form. Once the targets settle into a fixed directory layout with stable naming conventions, the judgment is encoded as a deterministic pipeline: embeddings + clustering + threshold logic, with at most a single bounded LLM call where unavoidable. Sits in the LLM Workflow Quadrant (Quadrant 3) or even the Script Quadrant (Quadrant 1) when the LLM call disappears entirely. Target identifiability is high; scale-resilience matters because the operation form runs over the full settled corpus rather than a per-invocation slice.
• Phase placement. The two forms sit in different lifecycle phases of the same project. Both can coexist; neither is a generalization of the other.
• Quadrant routing for each form.
  • Design phase form — Quadrant 4-leaning skill, but invoked inside a design session, so ADR-0010 (1) Phase-crossing decision is satisfied automatically (this is the design phase). The anti-patterns.md "Skill-internal phase descent ignored" entry guards against forcing this form into a frozen pipeline before the targets have settled.
  • Operation phase form — Quadrant 3 (or Quadrant 1) component in the default operation composition. No Phase-crossing decision is needed because the form is no longer Quadrant 4.
• Why capability is not the lever. A higher-capability LLM does not give the design-phase form fixed targets, and does not give the operation-phase form runtime tolerance for novel unseen variations. AKC's "capability ↑ → holistic judgment OK" principle does not cover either of the two secondary forces. Phase is what decides which form is structurally honest.
• Recursion inside a single skill. The same gradient applies inside a single skill: a design-phase skill can have one subcomponent that is a frozen lookup (target identifiability is locally high — for example, a config file at a known path) and another subcomponent that is runtime LLM judgment. The diagnostic frame walks recursively.
• Lineage note. The concrete examples behind this case are in the 2026-05-02 essay; they are not named in this navigator because the navigator is harness-neutral. The essay walks two AAP-author skills that exhibit the design-phase / operation-phase split for the same job.

A note on what these cases are not

The case studies above are illustrative routings, not endorsed reference architectures. The repository does not maintain a "reference architecture" file because reference architectures dissolve faster than the routing judgment behind them. Once the quadrant is identified, the architecture follows from the quadrant's properties (post-hoc separability, named gap-bearer commitment) and the applicable ADRs — not from a template.