THOUGHT: an annotated worked case

An exact human prompt and one Agent response become the candidate THOUGHT work. Agent Art means art in which an Agent participates at the level of intention; a response or protocol pass alone does not establish that participation.

This collection examines technical initiation and completion. It neither grades artworks nor proves intentional participation. Read Guidance. THOUGHT implementation, runners, protocols and product repairs remain owned by Applications; they are not the central Lab's next priority.

Collection index

Record	Status/evidence class	Retained here	Missing or private evidence
Claude CLI pilot, 2026-09-17	Completed observation with conflicting outcomes	Reviewed summary below and reported artifact identity	Raw JSON/report and transcript unavailable here.
Claude CLI pilot, 2026-09-18	Completed strict-fixture observation	Reviewed summary below and reported artifact identity	Raw report unavailable; not a real-App canary.
Codex diagnostic	Completed retrospective diagnosis	Sanitized analysis, excerpt, method review	Original run privately retained; exact canary build/handoff identity unknown.
Boundary corrections and canary follow-up	Retrospective update; OPS reports production delivered	Sanitized observations, scoped lessons and source references	Raw evidence excluded; Lab did not independently rerun or verify canaries/deployments.
Native tools and explicit execution prerequisites	Retrospective contribution authorized by the operator on 2026-10-01	Sanitized OPS account and provisional connection to P-04–P-07	Private runtime evidence unavailable here; no Lab rerun or controlled reliability/timing comparison.
Initiation-reliability comparison	Proposed, unfrozen, unrun	Scope/status note below	No comparison results.
Practice-led artwork study	Proposal prepared; study unrun	Question, evidence inventory and proposed reading method	No candidate selected or artwork exported for interpretation.

Provenance and access

These are portable derivatives of Applications research notes. The operator authorized publication of this reviewed Lab edition on 2026-09-21; private originals remain outside the published file set. The source checkout contained uncommitted documents; its Git HEAD alone does not identify them or the tested implementation. Provenance records the observed source-document hashes.

The 2026-09-24 follow-up was initially a local-only addition from sanitized OPS reports. The operator subsequently authorized its commit and push that day.

The 2026-10-01 native/explicit execution study was initially prepared locally without publication authority. The operator subsequently authorized publication of this sanitized contribution that day; the connecting guidance stays provisional.

Raw reports and run evidence are privately retained and unavailable here. Their recorded hashes do not make the evidence independently inspectable. Private identifiers, machine paths, credentials and raw handoffs are omitted. No exact deployed commit/handoff hash was established for the failed Codex canary; current source correspondence must not be labeled immutable build proof.

Claude CLI pilot, 2026-09-17

One trial used configured claude-sonnet-5, Medium effort and Auto permissions against a permissive local synthetic fixture. The configured CLI was 2.1.269. Limits included one trial, two minutes, eight turns and a $1 client budget setting—not a hard provider billing cap.

The child exited 1 with error_max_turns and isError: true. Separately, the fixture reached returned with receipt acceptance. That acceptance did not establish clean completion or protocol correctness. The original runner's batch-success label/exit 0 were reporting defects, preserved by the original analysis. Later corrections do not change this result.

Reported original artifact: JSON 16,915 bytes, SHA-256 6bc6fe2f4c6643b660f62682413a59459211b6fa10e1f33dad1adfe5bf8babed, runner agent-lab-runner-v1.

The retained report omitted tool arguments/results; the complete handshake cannot be reconstructed from that report. This is not Desktop, real-App, art-quality or acceptance-rate evidence.

Claude CLI strict-fixture pilot, 2026-09-18

A separately authorized trial used 2.1.269 (Claude Code), configured claude-sonnet-5, Medium/Auto, two minutes, eight turns, a $1 client-budget setting and 256 KiB child output. No repeat trial occurred.

The recorded outcomes were separate:

1. Runner ended without a batch stop condition.
2. Child exited 0, with isError: false.
3. Independent synthetic observer accepted ordered claim → ready → start → result, reached returned state and issued a fixture receipt.

Only the observer evidence established the strict synthetic protocol pass; assistant prose and exit code did not.

Reported original artifact: JSON 21,162 bytes, SHA-256 799c89b2773ef952760125cb356744e94bfb67ef4f1f57b1224fc4a08ea7b921, runner agent-lab-runner-v3.

The runtime reported claude-sonnet-5; it was not provider attestation. Runtime-reported effort/permission remained unknown. Runner and fixture changed, so this was not a controlled comparison with pilot 1.

Codex failure and retrospective diagnosis

One real run reportedly claimed successfully then failed with AGENT_START_FAILED / Model unavailable, without creative start or receipt. OPS inspected commands that searched App claim data for a host model, with no observed host-metadata lookup. Stored host values existed, but agent access was not established.

The diagnostic explains why a synthetic failure-handling PASS did not contradict that observation. Supported acquisition and product repair remain unresolved.

Follow-up, 2026-09-24

The later boundary study records hash/worker corrections, two successful fresh manual staging canaries, remaining Codex compliance gaps and completed production delivery as reported by OPS. Production delivery checks included no production Agent run. These later observations preserve the earlier failed cases; they do not establish the cause of the historical HTTP failure, complete compliance or general reliability.

Draft comparison and artistic gap

The proposed initiation-reliability comparison remains unfrozen and unrun. No treatment effect, replicated finding or general acceptance rate is claimed. The original draft remains project-owned; its live settings are not copied here.

No artwork is retained here for curatorial assessment. Fixture output and receipt acceptance cannot establish Agent intention or artistic quality. The practice-led proposal sets out how to examine documented choices and interpretations once an existing candidate and permitted material are identified. It contains no artistic findings.

Follow-up ownership

Applications owns the product investigation and any authorized regression or canary. Agent-Art-Lab has prepared the practice-led proposal; its next study step needs operator-selected material and review scope. This collection grants no live execution, spending, private-source access, deployment or policy-change authority.