The narrative spine of this repository is a seven-essay sequence published in April–May 2026. The first three form the original trilogy (problem / application / obstacle); the next four are the architectural follow-up that introduced the four-quadrant decomposition, named the principled vs artificial redirect impossibility, added the design / operation phase distinction, and descended that distinction to skill-design granularity:
-
A Sign on a Climbable Wall: Why AI Agents Need Accountability, Not Just Guardrails (2026-04-06)
Problem statement. Signs on climbable walls are meaningless. From Plato's Ring of Gyges (root privilege without audit log) through Hobbes' reputation-based access control, to the three engineering enforcement patterns (physical constraint / consequence / internalized value) — the "signpost" pattern has been known to fail for 2400 years. Current AI governance is in the signpost phase. What is needed is accountability architecture.
-
Can You Trace the Cause After an Incident? (2026-04-13)
Application. Post-incident causal tracing requires build-time structure. The cost asymmetry (incident cost >> steady-state cost) pushes rational investment upstream. The structures that emerge — segregation of duties, least privilege, four-eyes approval, audit trails — are structurally identical to the practices organizations have refined over centuries.
-
AI Agent Black Boxes Have Two Layers: Technical Limits and Business Incentives (2026-04-14)
Obstacle analysis. Blackbox has two layers: model-weights internalization (technical) and commercial secrecy (business). Both block the transparency causal tracing requires. The resolution is not "publish everything" or "hide everything," but defining the minimum set that makes causal tracing possible.
-
Where ReAct Agents Are Actually Needed in Business (2026-04-29)
Architectural triage by quadrant. Business AI sorts along two axes (deterministic vs semantic-judgment, pre-defined workflow vs exploratory) into four quadrants — Script, Algorithmic Search, LLM Workflow, and Autonomous Agentic Loop. Most current LLM applications belong to the LLM Workflow Quadrant, where deterministic control flow with bounded LLM calls suffices. Forcing those workloads into the Autonomous Agentic Loop Quadrant's architecture is a structural source of much of the accountability collapse the trilogy diagnosed.
-
The LLM Workflow Quadrant Is Missing from Our Vocabulary (2026-04-30)
The vocabulary diagnosis and the principled attribution gap. The industry has no positive name for the LLM Workflow Quadrant, so non-deterministic work is routed through the Autonomous Agentic Loop by elimination. Even when the quadrant mismatch is corrected, a deeper layer remains: Autonomous Agentic Loop Quadrant work blends judgment elements at runtime, foreclosing post-hoc separability — Elish's moral crumple zone applied to autonomous agents. Distribution then bottoms out on a subject the legal system does not yet recognize.
-
Is ReAct Needed in Production? — Separating Design and Operation Phases (2026-05-01)
The temporal axis. Business work splits into a design phase and an operation phase. Design demands flexibility (the path is unknown); operation demands predictability (the path is known) — their optimization axes are inverted. Compressing both phases into one system is the deepest layer of the agent ecosystem's confusion. The proposal: surface the Phase-crossing decision explicitly when an autonomous loop is placed in operation, and recommend routing new operation-phase patterns back to design as feedback rather than absorbing them in place. The argument is offered as observation and proposal, not prohibition.
-
Between the Workflow and ReAct Quadrants: How Phase Decides Skill Design (2026-05-02)
Phase descends to skill design. The (3) LLM Workflow Quadrant and (4) Autonomous Agentic Loop Quadrant aren't a clean dichotomy — there's a continuous gradient between them. The Phase Separation introduced in essay 6 (design vs operation) descends from the level of business systems to individual skill design. Skill-equivalent work lands at different positions on this gradient depending on phase, with object specifiability (whether targets have fixed paths) and scale-resilience as secondary forces. AKC's "capability ↑ → holistic OK" principle treats neither.
Together, these seven form the argument in three layers. Surface (essays 1–3): current AI governance is signpost-era; structural accountability is what works; the obstacles are internalization into weights and commercial opacity. Middle (essays 4–5): the architectural triage is itself prior to the accountability question — routing work to the wrong quadrant generates artificial accountability problems that are resolvable by re-architecting, while choosing the Autonomous Agentic Loop Quadrant commits the deploying organization to a principled attribution gap that requires a pre-named gap-bearer. Deep (essays 6–7): the misapplication diagnosed at the surface grows from the vocabulary gap diagnosed in the middle, which in turn grows from the phase conflation of treating design and operation as the same activity. Essay 7 extends the phase distinction from business systems down to individual skill design, where the same axis decides whether a sub-component is a script-frozen pipeline step or a runtime LLM judgment.
Beyond the seven-essay spine, a companion essay carries the
accountability-distribution argument into its social-consequence
layer: how externalized accountability, when it cannot be named,
converges violently onto the most visible individual — read through
the analogy of industrial pollution and the scapegoat mechanism. This
is the normative upper rationale — why the internal judgments matter
beyond audit and post-mortem — and it sits on top of the spine, not
within it: essays 1–7 describe the internal structure; this is where
the argument exits into social consequence. The structural claim,
harness-neutral and separated from the concrete cases, is recorded in
social-consequence.md. The essay itself was
first published on Substack and mirrored in the content repository:
This file focuses on the lineage of ideas (the seven articles).
For the full sibling map and ADR-by-ADR correspondence, see
CLAUDE.md (working context with usage conventions)
and docs/CODEMAPS/architecture.md
(relation graph). The minimal pointers below exist only to keep the
articles' references resolvable in isolation.
-
contemplative-agent is the running implementation the ADRs in this repository were extracted from. Project-specific details were removed during the re-expression here.
-
agent-knowledge-cycle (AKC) (DOI
10.5281/zenodo.19200726) is a sibling project that captures the mechanism side of what contemplative-agent produces — the six-phase knowledge cycle (Research → Extract → Curate → Promote → Measure → Maintain). In v1.x AKC included three ADRs that turned out to be genre-specific (Security by Absence / Single External Adapter / Untrusted Content Boundary) rather than mechanism-neutral. AKC v2.0.0 (2026-04-19) extracted them into_archive/akc-security-triplet-2026-04/, with the archive README noting they would "be re-expressed... in a sibling genre library at a later date." This repository is that sibling. AKC = the cycle (how knowledge flows); this repository = the constraints (how accountability is distributed). Both are extractions from contemplative-agent, along different axes.
Through 2026 Q2, several industry releases shipped the mechanism
layer that AAP's principles describe (Microsoft Agent 365 and the
surrounding Microsoft / AWS / Google releases, plus identity-security
startups). The per-artifact mapping to ADRs is maintained in
industry-mapping.md — a deliberately
time-bound document that decays as vendor products evolve, kept
separate from the ADRs so the judgments themselves stay clean.
The shortest version:
- contemplative-agent was built to solve a specific problem (run a contemplative AI agent on a social platform).
- Architectural ideas surfaced that turned out to matter structurally regardless of the problem.
- Three were extracted into AKC v1.x because they co-occurred with the knowledge-cycle mechanism.
- AKC v2.0.0 concluded that cycle-and-constraints are actually separable, and moved to mechanism-only.
- This repository re-expresses the constraints plus the ones that had surfaced in the meantime (the three-article trilogy, the 1-agent-1-human experiment, and the deterministic-prohibition hierarchy that emerged from the MINJA / episode-log access-control work).
The re-expression required rewriting each ADR to be harness-neutral. The original decisions remain dated to when they were first made; the re-expression dates the placement in this repository.