ADR-0010: Phase Separation Between Design and Operation (experimental)

May 2, 2026 · View on GitHub

Language: English | 日本語

ADR-0010: Phase Separation Between Design and Operation (experimental)

Summary. Phase (design vs operation) and Quadrant (problem-space) are independent dimensions of a deployment. Both phases can host components from any Quadrant, so this ADR is not a phase-to-quadrant mapping. The judgment it records is narrower: placing an Autonomous Agentic Loop Quadrant component as the operation-phase default route a runtime-driven attribution gap into the production environment, where ADR-0009's design-time commitment alone is no longer sufficient. Such placements remain admissible, but call for a documented Phase-crossing decision in addition to ADR-0009's gap-bearer naming.

Status

experimental — the design / operation distinction is well established outside the agent context (development vs production, as-is vs operation, design-time vs runtime), but its translation into the AAP Quadrants is recent (2026-05-01) and the operationally measurable signals for the operation-phase default are still being refined. See open questions at the end.

Date

2026-05-01 (first formalization in this repository)

Context

ADR-0009 (Triage Before Autonomy) answered "should this work adopt an autonomous-loop architecture at all?" by routing along the Business AI Quadrants. The 2026-05-01 essay surfaces a related but distinct question: if the answer is yes, where in the lifecycle does the autonomous loop live, and does that placement carry costs the design-time triage missed?

The essay's argument is offered as observation and proposal, not as a categorical mapping. The phrasing throughout — "isn't it really the case that…", "couldn't this be released…" — is diagnostic, not prescriptive. AAP picks up that posture: the rule below is narrow, the surrounding context is thinking-stimulus for the deploying team.

Phase and Quadrant are independent

The Phase / Quadrant landscape is two-dimensional — every Quadrant appears in both phases of real systems. See quadrants/README.md — Phase is a third dimension for the framing and orthogonality argument.

The 4×2 cross-product names typical placements:

	Design phase	Operation phase
Script Quadrant	prototype scripts	production pipelines
Algorithmic Search Quadrant	algorithm prototyping	production solvers
LLM Workflow Quadrant	prototype workflows in design reviews	production-default LLM workflows
Autonomous Agentic Loop Quadrant	coding agents, Deep Research, exploratory R&D	continuously-evolving research operations, exploratory analytics

The 4×2 matrix is a landscape, not a 1-to-1 mapping. The narrow judgment AAP records is not "Quadrant N is the X-phase tool." It is: the operation-phase placement of the Autonomous Agentic Loop Quadrant has a structural cost ADR-0009's design-time commitment alone does not cover.

Why operation-phase autonomous loops are the load-bearing case

ADR-0009's gap-bearer naming addresses the attribution gap as a deployment commitment. It works because ADR-0009 was framed at design time: at the moment of architectural choice, the deployer acknowledges the gap and names a subject who carries it.

Operation-phase placement adds a second-order risk on top: not just "the gap exists and someone bears it," but "the gap fires inside production runtime, against unanticipated patterns, while the gap-bearer is not in the loop." The 2026-05-01 essay's central observation: when a new pattern surfaces in operation, an autonomous-loop architecture absorbs it dynamically — and that dynamic absorption animates the runtime with the same blending that forecloses redirect. The artificial redirect impossibility of ADR-0009 reappears, this time inside the deployed system rather than as a design-time mismatch.

This is what the deploying team should think about, not blanket phase-to-quadrant mapping.

The three-layer diagnosis (essays 4–7)

Layer	Diagnosis	Essay
Surface	Quadrant misapplication — (3) work routed through (4) architecture	Essay 4 (2026-04-29)
Middle	Vocabulary gap — no positive name for the LLM Workflow Quadrant	Essay 5 (2026-04-30)
Deep	Phase conflation — "always-on autonomous agent runs business" compresses design + operation into one system; the same Phase axis descends to skill-design granularity (skill subcomponent level)	Essays 6 (2026-05-01) and 7 (2026-05-02)

Misapplication grows from vocabulary gap; vocabulary gap from phase conflation. Essay 7 extends the deep layer by showing that the same Phase axis applies inside skill design: a single skill can host both frozen-pipeline and runtime-judgment subcomponents, and the diagnostic frame applies recursively to each. ADR-0010 records the deepest layer as a diagnostic frame, not a partitioning rule.

Decision (experimental)

The Phase / Quadrant landscape is the deployment's design space. Within it, AAP records the following narrow judgment:

Operation-phase Autonomous Agentic Loop placements require an explicit Phase-crossing decision in addition to ADR-0009's gap-bearer naming. The Phase-crossing decision answers, in one sentence: when a new pattern surfaces in operation, will the autonomous loop handle it dynamically in place, or will it be routed back to the design phase as feedback? Both answers are admissible. "We'll figure it out when it happens" is not.
The default operation-phase composition is Quadrant 1 + 3 (+ 2 where applicable). This is a default, not a prohibition. It reflects an empirical observation: most workflows whose path is knowable at design time are better served by a composition that makes redirect succeed at the component level. Deviations from the default are admissible if (1) is satisfied.
Cross-phase use of any Quadrant is normal. The same Quadrant can appear in design and operation phases of the same project — a workflow prototyped (design) and then promoted (operation), or an autonomous loop used to explore (design) before its findings are encoded into a workflow (operation). The judgment is per placement, not per Quadrant.
Feedback path is the recommended response to new operation-phase patterns. When a new pattern surfaces in operation and (1) was decided as "route to design," the handoff is explicit: tag the pattern, escalate to design phase, and return the operation surface to the default composition until the design phase decides where the new pattern fits in the Quadrants. This is a recommendation, not a hard requirement; the alternative (in-place dynamic handling) is admissible under (1) and is explicitly chosen.

When these conditions are not named, the default in (2) applies: the operation deployment runs on Quadrant 1 + 3 (+ 2) components, and Autonomous Agentic Loop placements that drift into operation without an explicit Phase-crossing decision are out of compliance with ADR-0010.

Skill-design descent (informational, from essay 7)

The Phase axis above is stated at business-system granularity (deployment placements of Quadrants), but the same axis descends to skill-design granularity and even to skill-internal subcomponents. Essay 7 (2026-05-02) records the descent. It is informational here, not a load-bearing rule: the Phase-crossing decision in (1) is still required only at the business-system placement of Quadrant 4, not at each skill subcomponent.

The descent has three load-bearing observations:

The same job lands at different positions on the Quadrant 3 ↔ Quadrant 4 gradient depending on phase. A job whose target artifacts and execution environment are not yet known at design time naturally lives as a runtime-judgment skill (closer to Quadrant 4); the same job, once its targets settle into fixed paths and naming conventions in operation, can be re-implemented as a frozen pipeline of bounded calls (closer to Quadrant 3 or even Quadrant 1).
Two secondary forces decide where on the gradient a skill (or subcomponent) lands. Target identifiability asks whether the artifacts the skill operates on sit at fixed paths or under fixed naming conventions. Scale-resilience asks how many units the skill processes per cycle and how much miss rate is tolerable at that scale. Both must hold for the operation-phase frozen-pipeline form to work.
Capability is downstream of phase, not the primary lever. Higher-capability LLMs do not erase the differences in target identifiability or scale-resilience. AKC's Design Principle ("capability ↑ → holistic judgment OK") covers neither, so model choice is informed by the phase decision rather than driving it.

For the purpose of the Phase-crossing decision in (1), a skill is treated as a single business-system component. Internal phase gradients within a skill are governed by the skill-design gradient guidance, not by ADR-0010 (1). The anti-patterns.md and case-studies.md navigators operationalize the descent.

What this is not saying

It is not saying "Autonomous Agentic Loop is for design phase only." Operation-phase placements are admissible. They carry the additional commitment in (1).
It is not saying "LLM Workflow Quadrant is for operation phase only." Design-phase prototyping with bounded LLM calls is routine and legitimate.
It is not partitioning Quadrants by Phase. The Phase / Quadrant landscape is two-dimensional; this ADR records a single load-bearing observation about one cell of it (operation-phase Quadrant 4) and a default about another (operation-phase composition).
It is not banning anything. It is making the team's Phase-crossing decision explicit instead of implicit. The rule is procedural, not architectural.

What this permits

The same workflow appearing in both phases. A workflow whose routing prompt is prototyped in design and promoted to operation is the normal lifecycle, not a Phase boundary violation.
Operation-phase Autonomous Agentic Loop with the Phase-crossing decision recorded. This is admissible. The deployment is accountable for the cost it has named.
Design-phase residents that look like deployed services. Coding agents, Deep Research tools, exploratory browsing agents — these run continuously, but each session is an independent design session that completes when its exploration finishes. Their continuous availability is a deployment-level fact, not a phase-level property of each session.

What this does not permit

Implicit Phase-crossing decisions. Operation-phase Autonomous Agentic Loop placements without a documented Phase-crossing decision recorded at deployment time are out of compliance.
Operation-phase autonomous loops without an ADR-0009 gap-bearer. ADR-0009 already requires this; ADR-0010 does not relax it.

Alternatives Considered

Strict ban: autonomous loop never enters operation phase. Rejected. Records and observations show legitimate edge cases (high-frequency exploratory analytics, continuously-evolving research operations) where the design / operation boundary blurs. A categorical ban writes those cases out of the framework rather than naming the cost they pay.
Phase-to-Quadrant 1-to-1 mapping (Autonomous Loop = Design phase, LLM Workflow = Operation phase). Rejected as the earlier draft of this ADR did. Real systems put Quadrant N in both phases routinely; a 1-to-1 mapping forces them out of the framework. The mapping is a landscape, not a partition.
Strong guardrails are sufficient. Rejected as a substitute for the Phase-crossing decision. Guardrails (sandbox strength, HITL load) bound blast radius and add review; they do not surface the team's choice between in-place absorption and feedback-to-design as an explicit deployment-time decision.
Subsume Phase Separation into ADR-0009. Rejected. ADR-0009 triages whether the work belongs in the Autonomous Agentic Loop Quadrant at all. ADR-0010 surfaces the Phase-crossing decision for placements where ADR-0009's answer is yes. They are independent triage axes (problem-space and time-axis); merging them loses both clarity and the diagnostic-frame value of the three-layer analysis.

Consequences

The repository's structure becomes 10 ADRs + 4 Business AI Quadrants, where ADR-0009 and ADR-0010 form a triage pair: 0009 asks "which Quadrant?", 0010 surfaces the Phase-crossing decision when the answer is the Autonomous Agentic Loop Quadrant.
Coding agents, Deep Research, and similar always-on exploratory systems are explicitly named as design-phase residents deployed as services. Their legitimacy under AAP is preserved without forcing them into a category exception.
Operation-phase deployments default to a transparent composition (Quadrant 1 + 3 + 2 where applicable), where redirect succeeds at the component level. Deviations are admissible and explicit.
Design-phase use of LLM Workflow Quadrant or Autonomous Agentic Loop Quadrant is normal and not flagged.
Post-incident reconstruction can answer not only "what did the agent do?" but also "what Phase-crossing decision was the deployment running under when this happened?" The answer is in the deployment record, not inferred.
The three-layer diagnosis (essays 4–7) becomes a permanent diagnostic frame for design reviews: when an autonomous loop is proposed for operation, walk the layers. Most resolutions surface a misapplication (layer 1) or a vocabulary gap (layer 2) before the phase question even arises.
The same diagnostic frame is now usable at skill-design granularity: a single skill, or a single subcomponent within a skill, can be walked through misapplication / vocabulary gap / phase conflation in the same order. This descent is informational; ADR-0010 (1) still applies only at the business-system level.

Open questions (why this is experimental)

Operation-phase Autonomous Agentic Loop edge cases. High-frequency exploratory analytics, continuously-evolving research operations, real-time content moderation against unbounded patterns — these are domains where the Phase-crossing decision "handle in place" may be the structurally honest answer. AAP does not have empirical signals yet for distinguishing "Phase-crossing in place is right here" from "this team is rationalizing a misapplication."
Sequence-of-design-sessions framing. When does a coding agent or Deep Research tool stop being a sequence of independent design sessions and become an operation-phase deployment? The structural criterion (each session completes; state does not accumulate across them in operation-relevant ways) is open to interpretation in vendor-built systems.
Recursion. When the design phase itself is autonomous (the agent designs its own next experiment), does the rule recurse? The boundary between "design session" and "design system" is unresolved.
Interaction with ADR-0009 in operation-phase placements. Does the Phase-crossing decision constitute a second commitment alongside the gap-bearer (one for the attribution gap, one for the phase-crossing risk), or is the same subject both? In practice, the same human / role is usually both; whether a structural separation is needed is open.
Vendor framing. Most current autonomous-loop products are marketed as operation-phase deployments. AAP-aligned deployments need to make the Phase-crossing decision explicit even when the vendor materials assume "operation = autonomous-loop runtime" by default. Whether this requires a counter-vocabulary AAP should provide is open.
Skill subcomponent descent and the Phase-crossing decision. Essay 7 surfaces that the Phase axis descends inside individual skills. AAP currently treats a skill as a single business-system component for the purpose of (1), but the boundary between "skill subcomponent that legitimately runs as runtime judgment" and "skill subcomponent that smuggles a Quadrant 4 placement into operation phase without a Phase-crossing decision" is under-specified. Whether a future revision should require a per-skill Phase-crossing decision when any subcomponent is runtime-judgment is open.

This ADR will be revised as operational experience accumulates with the Phase-crossing decision in actual deployments. The status will move from experimental to accepted once the edge cases have documented answers, or to rejected / superseded if the design / operation distinction turns out not to be the right cut.

Lineage. Articles 6–7 of the seven-essay narrative spine (2026-05-01 / 2026-05-02) jointly form the lineage. Article 6 articulated the design / operation distinction and the three-layer diagnosis (misapplication → vocabulary gap → phase conflation) at business-system granularity. Article 7 descended the same Phase axis to skill-design and skill-subcomponent granularity, observing that the Quadrant 3 ↔ Quadrant 4 boundary is a continuous gradient on which the same job takes different shapes depending on phase, with target identifiability and scale-resilience as secondary forces. The argument across both essays is offered as observation and proposal, not as a categorical phase-to-quadrant mapping; ADR-0010 preserves that posture and absorbs essay 7 as informational descent rather than as a new load-bearing rule (the (1) Phase-crossing decision still applies only at business-system placements). Parallel to ADR-0009: 0009 triages whether the work adopts the autonomous loop; 0010 surfaces the Phase-crossing decision when 0009's answer is yes and the placement is operation-phase. The design / operation distinction exists outside the agent context (development vs production, as-is vs operation, design-time vs runtime); the agent context imported the cut rather than inventing it.