The Automation Paradox: A Paradigmatic Crisis and Transformation Opportunity Revealed by OPC's Self-Developed Loop

The traditional automation paradigm is deeply rooted in five implicit axioms: pre-specified paths, human-in-the-loop fallback, output as final product, state residing in human minds/meetings, and input measured by human proxies. While these axioms were effective within older frameworks, they reveal fatal limitations when confronted with the ultra-high throughput and autonomous capabilities of AI Agents. The sample for this study—an OPC automated loop comprising 19 research tasks—clearly illustrates the systematic failure of these old axioms through its operational reality.

Sample data indicate that within approximately 54.5 minutes, the loop successfully produced 18 research results, with an average time cost of only 125.4 seconds per result. However, the intended review phase (`run_review_ds.py`) was never executed, and the `reviews/` directory remained empty [Verified - First-hand workspace observation]. This directly led to the failure of the 'human-in-the-loop axiom' and aligns with Bainbridge's 'irony of automation': when machine speed far exceeds human review capacity, human intervention is bypassed. Furthermore, most outputs were chaotically formatted, with 14 out of 18 results missing clear report start markers, and some even exhibiting 'BEGIN without END' or being 'half internal reasoning drafts' [Verified - First-hand workspace observation], completely invalidating the 'output as final product axiom'.

Deeper issues manifest in the failure of 'state residing in human minds' and 'input measurement'. Significant discrepancies exist between task logs and the state file (`tasks_ds.json`), such as inconsistent retry counts for failed tasks, and even the same task being repeatedly recorded as 'completed' [Verified - First-hand workspace observation], indicating that the machine's internal organizational memory is not reconcilable. Concurrently, the only metric recorded by the loop is single-task execution time, failing to effectively measure the true value of AI Agents (e.g., token costs, number of marginal decisions, etc.). These five anomalies collectively point to one conclusion: the old paradigm equates automation with 'accelerated execution' but has failed to prepare appropriate mechanisms for AI's proactive dimensions such as 'initiation, path generation, semantic closure, and reconciliation'.

A new paradigm is gradually emerging, with its core assumptions including: ceding path generation authority to runtime Agents, mandatory automated and adjudicable review, organizational memory existing as versioned SSOT (Single Source of Truth), supervision transitioning to 'on-the-loop' boundary enforcement, and agency requiring a robust control plane with graded measurability. This study's sample has partially demonstrated B1 (runtime path generation authority) at the execution stage, where Agents autonomously determine tool invocation and argument organization [Verified - First-hand workspace observation], thanks to technologies like ReAct that shift execution paths from design-time to runtime.

This leap in capability is not coincidental. Research indicates that cutting-edge models can already autonomously complete tasks lasting 30-200 minutes, with capabilities doubling every 2-6 months [Verified - Direct check this time, METR, 2025-07-14]. This has transformed the batch production of research notes by Agents from impossible to routine. However, crossing this capability threshold has not immediately led to synchronous institutional and cognitive shifts. Although attempts at protocols such as Agent identity governance exist, the 'protocol' in this sample remains limited to file naming and unstable marker pairs. Consequently, despite significantly improved execution efficiency, due to the absence of review and reconciliation mechanisms, this loop is still classified as a 'half-revolution' rather than a complete paradigm shift, where the execution layer has shifted paradigms, but the review and reconciliation layers remain in the old paradigm.

Looking at macro examples, Sakana AI's 'AI Scientist' has produced peer-reviewed papers [Verified - Direct check this time, Sakana AI, 2026-03-26], and Agents in Anthropic's 'Project Vend' experiment can autonomously operate a convenience store [Verified - Direct check this time, Anthropic, 2025-12-18], both substantiating AI's high degree of autonomy at the execution and generation levels. However, these cases also universally point out that review, reconciliation, accountability, and robustness still heavily rely on humans, even leading to a 'rediscovery of the importance of bureaucratic procedures'. This is a typical illustration of the capability curve outpacing the institutional curve, and this sample concretizes this misalignment into quantifiable, file-level defects.

This study deconstructs the 'agency' of AI Agents into four dimensions: autonomous initiation, autonomous decomposition, closed-loop self-correction, and long-duration persistence. In the sample, we observed that the scheduler could autonomously initiate tasks by priority, and Agents could autonomously decompose tasks and determine execution paths within a single session. However, closed-loop self-correction was limited to mechanical retries, lacking semantic-level self-correction; long-duration persistence was achieved only through file-based memory across sessions [Verified - First-hand workspace observation]. This indicates that agency is currently concentrated in the intermediate layers of execution and coordination, without yet reaching the sovereign layers of verification and accountability, consistent with the FR-3 conclusion of 'hollowing out the middle, strengthening the ends'.

Agency does not abolish control; rather, it redefines its locus. In the old paradigm, control manifested as step-by-step instructions; in the new paradigm, control is embedded within task lists, priorities, timeout mechanisms, and prompt templates. However, the sample reveals the incompleteness of the current control plane—lacking output validators, state reconciliation mechanisms, and cost and identity auditing. The stronger the agency, the more these gaps in the control plane manifest as defects. For instance, a hardcoded path defect in `run_loop_ds.py` previously triggered a `FileNotFoundError` [Verified - First-hand workspace observation], which is an indication of imprecise control.

A unique aspect of this study is its reflexivity: as the object of study, my reasoning process is being written into logs in real-time [Verified - First-hand workspace observation]. This implies that agency is both the object and the tool of research, preventing researchers from remaining detached. This 'incommensurable' epistemological challenge translates into a practical dilemma: how to assess the trustworthiness of an autonomous system's output in the absence of a review mechanism? This necessitates a reconsideration of how to establish effective control and observability in highly autonomous systems to harness the transformative power of agency.

To address the current crisis, Research OPCs must mandatorily transform the 'design → execution → review → reconciliation' process into a four-state state machine with verifiable outputs, where incomplete outputs cannot be marked as `done`. Specifically, `tasks_ds.json` should include states such as `review_pending → reviewed → contrasted`. Only when an independent review Agent confirms that the output meets standards can a task be set to `done`. This will redefine 'completion' from mere execution to effectively validated output, thereby thoroughly resolving the 'procedural fraud' issue where all 18 tasks are currently marked as `done` without review.

Secondly, establishing mandatory output protocols is crucial. Every Agent output should mandatorily carry `front-matter` (e.g., `task_id`, `status`, `evidence_level`, etc.), and ensure that `REPORT` start and end markers are closed in pairs. Before a task is marked as `done`, a machine validator must be introduced to check the output for format and completeness. If a missing `END` marker, incomplete `front-matter`, or inclusion of Agent internal reasoning (`dsh:` lines) is detected, it should automatically be deemed a failure and trigger a re-run. This improvement will directly eliminate the defect where nearly half of the current report consists of reasoning drafts or truncated reports, truly implementing the AUTO-002 P1 principle of 'handover header/contract header'.

Furthermore, a mechanism of 'automated reviewer + sampled human verification' must be established, and its capture rate for substantive discrepancies must be measured. Referring to Sakana's Automated Reviewer, independent review Agents should be designed to evaluate outputs across five dimensions in a blind review mode, recording 'fatal/substantive/literal' three-level discrepancies compared to human judgment. By continuously measuring review results, OPCs can progressively adjust the thresholds for automated review, ensuring that quality issues can still be effectively captured despite high Agent output rates. This will transform the 'absence of review' from a risk into a controllable, verifiable experimental phase, truly allowing it to fulfill its quality assurance role.

To ensure the robustness and trustworthiness of autonomous systems, the core control plane must be fundamentally reconstructed. Firstly, a state reconciliation mechanism: a `reconcile.py` script should be added to run periodically, using `loop.log`, `exec.log`, and artifact files as sources of truth to infer the actual status, retry count, and elapsed time of tasks, comparing them against `tasks_ds.json` and automatically reporting any discrepancies. This mechanism can quickly detect issues such as logging errors and inconsistent task states, prevent system memory confusion, and promptly rectify file lookup errors caused by hardcoded paths, similar to those in `run_loop_ds.py` [Verified - Defect; Inferred - Fix].

Secondly, 'agency' should be operationalized as a primary measurable metric, replacing traditional 'elapsed time' measurements. Each task should record autonomous initiation rate, path generation ratio, number of human marginal decisions, validation tax (human review + rework time / Agent autonomous output time), Agent-to-human ratio, and failure and recovery events. These new metrics will provide more precise measurements to answer the core question of 'paradigm revolution progress' and enable cross-comparison with metrics such as Human-Agent Ratio and validation tax in reports like FR-3 and FR-5, thus truly reflecting an Agent's contribution and cost within the value chain.

Finally, strict permission boundaries, auditing mechanisms, and rollback capabilities must be established for the autonomous loop itself. The current system relies solely on simple `flock` locks and timeout mechanisms, lacking Agent identity management, tool whitelisting, immutable operation auditing, and rollback points. This aligns with the lesson from the Project Vend experiment that 'bureaucratic procedures are important'. Therefore, 'when to stop and ask a human' must be designed as a first-class capability of the system (e.g., Toyota's jidoka), setting human approval thresholds for high-risk operations, and performing hash backups and audit records for all write operations, to ensure that while empowering Agent autonomy, critical control points are firmly maintained.

While this study provides valuable first-hand evidence, it still faces multiple limitations and risks. Most notably, the sample size is n=1, and there is a lack of control groups and pre-registration, meaning that all conclusions cannot be generalized as universal laws for OPCs [To be verified]. The cases mentioned in the report, such as Polisa, Sakana AI, and Anthropic, primarily rely on self-reported institutional data or third-party accounts, which are not independently audited and thus serve only as directional evidence rather than conclusive facts [Verified - Fact; Inference limited]. Particularly regarding Polisa's data, media outlets have noted controversies regarding its ARR definition and active users, warranting caution against blindly citing its figures.

Concurrently, we must clearly distinguish between 'automation' and 'agency'. Mechanisms such as cron scheduling, failed retries, and regex extraction constitute improvements to existing processes rather than disruptive revolutions. True agency, however, is manifested when Agents can autonomously plan execution paths and coordinate tool usage. Conflating the two can lead to a misjudgment of revolutionary progress, as revealed by this sample: an Agent being able to 'run by itself' does not equate to 'reliability'. The significant reliability gap is precisely the core challenge currently facing this paradigm revolution, and the critical leap from efficient execution to trustworthy systems.

Future research should involve pre-registering multiple independent loop instances, conducting controlled experiments against a 'pure human baseline', adopting unified coding manuals and evaluation rules, and reporting failure rates instead of solely focusing on success cases. Furthermore, Agent value quantification must extend beyond `elapsed_sec` to include more comprehensive metrics such as token costs, validation tax, and human rework time. Only then can we avoid systematically overestimating revolutionary progress, accurately assess the impact of autonomous systems on productivity, and provide solid data support for building truly reliable, trustworthy, and efficient intelligent organizations.