Language: English | 日本語
Summary. Before adopting an autonomous-loop architecture (one in which the agent decides each next step at runtime), triage the work along two axes — deterministic vs semantic-judgment, and pre-defined-workflow vs exploratory. Reserve the autonomous loop for the quadrant that genuinely requires open-ended exploration. Choosing the autonomous loop commits the deploying organization to absorb a non-removable attribution gap as a named, pre-deployment cost — not a post-incident discovery.
experimental — the principle follows structurally from the trilogy of essays (2026-04-06 / 04-13 / 04-14) and the architectural follow-ups (2026-04-29 / 04-30), but the operational signals for distinguishing post-hoc separable architectures from blended ones at design time are still being refined. See open questions at the end.
2026-04-30 (first formalization in this repository)
The trilogy and the eight earlier ADRs answer "what should be
constrained, and who is responsible." Each answer assumes a structural
premise that is rarely stated: each judgment element in the system is
post-hoc separable — when something goes wrong, the operator can
trace the failure to a specific component (a model selection, a
routing rule, an upstream data source) and route the responsibility
there. This is what makes accountability distribution
operate. Distribution requires a subject who can be redirected to.
The ReAct-style autonomous loop violates that premise. In an
autonomous-loop architecture, each iteration's role is decided at
runtime — the model output, tool selection, history reference, and
prompt context blend into a single stream of judgments. When the
output is wrong, the contributions cannot be untangled after the
fact. Attribution gap is the name
for this foreclosed separability. It is intrinsic to autonomy, not a
maturity problem that better tooling will solve.
The 2026-04-29 essay introduced a four-quadrant decomposition of business AI work along two axes:
| Pre-defined workflow | Exploratory | |
|---|---|---|
| Deterministic | Script Quadrant | Algorithmic Search Quadrant |
| Semantic-judgment | LLM Workflow Quadrant | Autonomous Agentic Loop Quadrant |
The 2026-04-30 essay observed that the industry's standard vocabulary has no positive name for the LLM Workflow Quadrant. "Anything that isn't deterministic" gets routed to the Autonomous Agentic Loop Quadrant by elimination. The result is two distinct redirect failure modes that have been confused with each other:
- Artificial redirect impossibility. LLM Workflow Quadrant work routed through autonomous-loop architecture. Resolvable by re-architecting.
- Principled redirect impossibility. Autonomous Agentic Loop Quadrant work, where blending genuinely forecloses redirect. Not resolvable architecturally.
Most current accountability discussion (sandbox strength, HITL overload, moral crumple zone — Elish 2019) addresses the artificial case. The principled case — what to do when a task genuinely requires open-ended exploration and attribution must still distribute — has barely entered the discourse.
The earlier ADRs cannot resolve this on their own, because the architectural choice that creates the gap is upstream of every constraint they describe. A sandbox bounds blast radius (ADR-0004); it does not restore redirect-ability. An approval gate gates external side effects (ADR-0005); it does not separate blended judgments. The triage decision belongs at design time, before the rest of the ADRs become applicable.
Before adopting an autonomous-loop architecture, triage the work along the two axes of the four quadrants:
- Script Quadrant (deterministic × pre-defined).
- Algorithmic Search Quadrant (deterministic × exploratory).
- LLM Workflow Quadrant (semantic-judgment × pre-defined). The documented default for most current LLM applications. Load-bearing property: the path is decided in advance, the LLM is called as a single bounded step within that path. Two sub-forms by I/O modality: (3a) Conversational (specialized chat agents, where the human in the conversation is the judging agent) and (3b) Batch (single-purpose LLM functions inside deterministic pipelines, where the pipeline owns the control flow).
- Autonomous Agentic Loop Quadrant (semantic-judgment × exploratory). Use only when the work itself requires open-ended exploration — bounding the LLM call's role would prevent the work from being done at all.
See quadrants/README.md
for the operational signal of each quadrant, and
quadrants/decision-tree.md for the
five-question triage that operationalizes this Decision.
Choosing the Autonomous Agentic Loop Quadrant commits the deploying organization to absorb the attribution gap as a named, pre-deployment cost. The named cost has two components:
- An organizationally identifiable subject who carries the gap. In current practice, this is a human (consistent with ADR-0008 One Agent, One Human). It cannot be the agent itself; agents are not yet legal subjects. It cannot be resolved post-incident; the moral crumple zone literature shows the burden then falls on the operator with the least actual control.
- An explicit acknowledgement in the deployment record that redirect for this work will not succeed at the level of separable contributions. Failure analysis will identify that the system failed; it will not produce a clean redirect target.
The triage decision is made at design time, by humans, and recorded alongside the architectural choice.
- Hybrid architectures that combine quadrants are compatible if the boundary between them is itself a structural choice and the Autonomous Agentic Loop portion is treated as such (with a named gap-bearer). Example: clause extraction in the LLM Workflow Quadrant, escalating to an approval-gated Autonomous Agentic Loop for novel-clause analysis.
- Moving work between quadrants as the work matures. Tasks that start as exploratory often become structured once the path is understood; re-architecting from Autonomous Agentic Loop to LLM Workflow is a legitimate maturation, not a regression.
- Routing decisions made by LLMs within the LLM Workflow Quadrant (Anthropic's routing pattern). Routing is a single bounded judgment with a documented schema; it does not blend judgments at runtime.
- Defaulting to the Autonomous Agentic Loop for work that is semantic but pre-defined. The category error this repository names.
- Deferring the triage to runtime. The decision to enter autonomy is itself an architectural commitment; letting the agent decide when to escalate inherits the same gap that motivated the triage.
- Naming the gap-bearer post-incident. If no organizationally identifiable subject was named at deployment time, the work is in the Autonomous Agentic Loop Quadrant without a gap-bearer — the configuration this ADR exists to prevent.
- Treating sandbox strength as a substitute for triage. Sandboxing bounds blast radius (ADR-0004 territory). It does not restore separable contribution.
- Adopt autonomous-loop architecture by default; constrain via guardrails. Rejected. Guardrails are signposts in the sense of ADR-0001's critique. The architectural choice itself foreclosed redirect; no guardrail restores it.
- Sandbox the autonomy strongly enough to make it safe. Rejected as a substitute. Sandboxing addresses blast radius (ADR-0004), not redirect-ability. A perfectly sandboxed autonomous loop still has an attribution gap.
- Defer the triage to runtime; let the agent decide when to escalate out of the loop. Rejected. Runtime triage is an autonomous decision and inherits the same gap. The decision belongs at design time.
- Treat the four quadrants as agent categories rather than work categories. Rejected. The quadrants describe the work, not the agent. The same agent process can be deployed against different quadrants of work; what changes is the architecture chosen for that work.
- Skip the triage, rely on ADR-0005 (approval gate) and ADR-0008 (one operator) to catch the consequences. Rejected. Both ADRs become more load-bearing under autonomy, not less. They cannot recover separability that the architectural choice has already destroyed.
- The LLM Workflow Quadrant becomes the documented default for most current LLM applications. The vocabulary normalizes that default rather than treating it as the deviation from "real" agents.
- When the Autonomous Agentic Loop is chosen, the cost is paid upfront. ADR-0005 (approval gate), ADR-0006 (causal traceability), and ADR-0008 (one operator) become more load-bearing under autonomy, not less.
- Designers must distinguish post-hoc separable architectures from blended ones at design time. Practical signals: each LLM call has a documented role; each call has an input/output schema; failures map to a specific component. If those signals do not hold, the architecture is in the Autonomous Agentic Loop Quadrant whether or not it was named that way.
- The four-quadrant decomposition becomes a permanent diagnostic frame paired with the ten ADRs — together forming a two-axis structure: the ADRs answer per-question, the quadrants route the work to where those answers apply.
- Vendors and reference architectures that publish autonomous-loop templates without naming the attribution-gap commitment are doing what ADR-0001 calls signposts on a climbable wall — the pattern this repository is designed against.
- Boundary cases. Multi-turn chat agents with bounded system-prompt-defined roles — separable enough to count as LLM Workflow Quadrant, or already in Autonomous Agentic Loop territory? The answer depends on whether the role bounds hold under prompt injection and unusual user input; criteria are not yet sharp.
- Operationally measurable proxies for post-hoc separability. "Each call's contribution is separable" is currently a design judgment. Empirical signals (call-graph stability across runs, consistency of role under perturbation, schema compliance rate) may make the judgment more transferable, but the right metric set is open.
- Interaction with ADR-0008 when no responsibility subject exists. If the work genuinely requires the Autonomous Agentic Loop Quadrant and no organizationally identifiable subject can bear the gap (high-stakes decisions outside any individual's authority — large-scale legal, medical, fiscal action), is the right decision to not deploy, to deploy with a contractual party (insurance pool, regulatory body) as the bearer, or to wait for institutional infrastructure that does not yet exist? This is manifesto territory.
- Interaction with regulatory regimes. Some domains will regulate the Autonomous Agentic Loop Quadrant directly (the EU AI Act's high-risk classification is a near-precedent). Whether the triage decision should reference such regimes by name, or remain architecturally framed and let regulation map onto the quadrants externally, is unresolved.
This ADR will be revised as operational experience accumulates with the Quadrants vocabulary in actual deployments. The status will move from experimental to accepted once the boundary cases have documented answers, or to rejected / superseded if the triage frame turns out to be the wrong frame.
ADR-0009 sits before ADRs 0001–0008 in the order of application — it
triages which of the eight become load-bearing for a given piece of
work. See
quadrants/governance-mapping.md
for the per-quadrant ADR application table. Summary: out of scope in
the Script and Algorithmic Search Quadrants; ADRs 0001–0007 (+ 0008
when there are external side effects) in the LLM Workflow Quadrant;
all nine in the Autonomous Agentic Loop Quadrant, with ADR-0005,
ADR-0006, ADR-0008 carrying the most weight.
This is why ADR-0009 belongs in the same document set as ADRs 0001–0008: not as a successor, but as a triage that determines which of the eight become load-bearing for a given piece of work.
Lineage. Articles 4 and 5 of the
seven-essay narrative spine: Article 4 (2026-04-29)
introduced the four-quadrant framing of business AI work; Article 5
(2026-04-30) named the principled vs artificial redirect impossibility
and the absent accountability subject (moral crumple zone, Elish
2019). The source essays' provisional names ("ReAct Quadrant",
"Classical AI Quadrant") are preserved as aliases in
../glossary.md; this ADR uses the renamed forms
to keep the Decision body harness-neutral. Parallel to ADR-0008
(One Agent, One Human): 0008 binds the
accountable human; 0009 binds the architectural choice that makes
0008's binding meaningful.