From Passive to Proactive: Why This Is the Watershed

**Passive instrument:**

1. No endogenous goal: the goal is always given externally

2. Pre-set path: the action sequence is hard-coded at design time

3. No initiation: it stays still unless triggered

4. No closed loop: it ends when execution ends — no checking, no retrying

**Proactive agent:**

1. Autonomous initiation: starts without step-by-step human instruction

2. Autonomous decomposition: the path is generated at runtime, not enumerated at design time

3. Closed-loop self-correction: reads feedback, revises the plan, retries, rolls back

4. Long-horizon persistence: maintains the goal without drift over dozens of steps or hours

【verified】 Anthropic's *Building Effective Agents* (2024-12-19) gives the industry's clearest dividing line: a **workflow** is "LLMs and tools orchestrated through predefined code paths"; an **agent** is "LLMs dynamically directing their own processes and tool usage."

As early as 1978, Sheridan & Verplank proposed "levels of automation"; Parasuraman & Riley extended them in 1997, with the highest level even including "computer decides everything and acts autonomously." 【verified·classic literature】

But those levels describe **engineering-automatable automation within closed loops** — goals, boundaries, and evaluation criteria are all preset by humans. True "proactivity" is not a higher level of automation; it is the **transfer of path-generation authority** from "hard-coded at design time" to "generated at runtime."

This is the essential difference: automation makes a system run faster along the human blueprint; proactivity lets a system draw its own blueprint under the human's goal.

The old world rested on five "self-evident" axioms:

| Axiom | Content | Point of failure |

|---|---|---|

| A1 Single agent | Humans alone hold goals and initiate action | Agents initiate autonomously for the first time |

| A2 Pre-set path | Automation boundary = enumerable boundary | The path becomes model output |

| A3 Human-in-loop | Human exit = loss of control | Autonomous closed-loop correction appears |

| A4 Effort accounting | Value measured in headcount/hours/lines | The definition of output is rewritten |

| A5 Closed tasks | Tasks decompose into closed stages | End-to-end autonomous delivery appears |

**Anomaly 1: The "last mile" of improvement-type automation can never be automated — because paths cannot be exhausted.** Sixty years of automation (scripts, RPA, rule engines, CI/CD, Kubernetes self-healing) all required "write the steps down first." Google SRE's toil paradox is the proof: toil grows linearly with scale and is never finished 【inference】.

**Anomaly 2: The stronger the tool, the more fixed the bottleneck.** METR 2025 randomized controlled trial: with AI allowed, task time increased 19% while developers self-reported 20% speedup; DORA 2024: every 25% increase in AI adoption reduced delivery throughput by 1.5% 【verified】. The old ruler cannot measure the new object.

The watershed is not "machines can do things" — it is that **the answers to the four questions can now, for the first time, be partly "agent"**:

| The four questions | Old paradigm | New paradigm |

|---|---|---|

| Who generates the path | Humans / human-written code | Agent (generated at runtime) |

| Who initiates action | Humans | Agent (autonomous initiation) |

| Who closes the loop | Humans | Agent (self-correction/retry/rollback) |

| Who bears responsibility | Humans | **Still humans** |

The first three questions can now be partly answered "agent"; the fourth must still be answered "human."

**Proactivity is not a matter of degree — it is the transfer of path-generation authority from design time to runtime.** When path-generation authority transfers, the bottleneck shifts too: from "execution capability" to "intent-definition quality" and "verification capability."

**Individuals**: The capacity boundary shifts from "personal time × energy" to "goal-definition quality × agent-cluster scale." The two scarcest capabilities become: defining good goals, and verifying good outcomes.

**Organizations**: The minimum unit shifts from "person + tool" to "person + agent cluster." Management cost collapses — because agents bear execution and closed loops, management shrinks to "goal governance."

**Business**: Value capture shifts from "human hours" to "replicable proactive systems." The definition of asset begins to point to "systems that advance themselves" — data, models, processes, verification loops.

1. **Don't treat agents as automation**: if you are still hard-coding every step, you are still in the old paradigm — just running faster.

2. **Move attention to "goal" and "verification"**: these two are the only non-delegable things.

3. **Build your own verification loop**: make the quality of every output measurable, traceable, and accumulating — this is the core asset of the new paradigm.

4. **Research OPC is a living sample**: its automated loop (design → execution → review → comparison) is the minimal real-world implementation of "proactivity."

The shift from passive to proactive is not a small step in tool evolution — it is a **transfer of agency ownership**. Understanding this watershed is how we understand why the paradigm revolution is not "a faster old world," but a different world.