Abstract

The common mistake in AI transformation is to treat work as a chain of commands. That model assumes a boss, a subordinate, and a clear top-down order of action. It is useful for some coordination problems, but it is too simple for human-agent systems. In practice, work is increasingly organized through environments, signals, traces, permissions, and local roles. The result is not a neat command hierarchy. It is a distributed system in which people and agents respond to the same inputs, operate under the same constraints, and produce coordinated behavior without a single commander controlling every move.

This essay argues that Reworking Order should think about work through an environment-and-signal model. The useful question is not whether an agent or a human is “in charge” in the abstract. The useful question is who initiates change, who authorizes it, what signals shape behavior, and what traces let the system be audited and improved. That also leads to a more practical operating principle: for most consequential change today, humans should initiate and authorize; AI should expand the option space; and over time, some lower-risk work may move along a continuum toward agent initiation with human review. The point is not to freeze the boundary. The point is to design it deliberately.

From Command To Signal

The classic manager model assumes that work is a sequence of instructions. Someone decides, someone executes, someone reports back. That frame still matters in many organizations, but it misses how modern work actually behaves once software, prompts, policies, automations, shared documents, and AI systems are part of the loop.

In a human-agent environment, the important unit is not the command. It is the signal.

A signal can be a goal, a policy, a prompt, a constraint, a template, a notification, a dashboard, a metric, a review comment, or a hidden expectation in the environment. Signals tell actors what kind of behavior is appropriate. They do not need to specify every step. In fact, if the system is healthy, they usually should not. The better-designed the environment, the less direct commanding it requires.

That is why the old boss-subordinate model breaks down. A human may still own the objective, but the actual movement of work is shaped by:

  • the interface
  • the permissions model
  • the prompt library
  • the review rule
  • the shared metrics
  • the visible history of prior decisions
  • the trace left by prior action

Once those pieces exist, behavior becomes partly emergent. Agents respond to the context they are placed in. Humans respond to the same context. Coordination comes from shared structure rather than constant instruction.

That is the intuition behind the “standalone complex” idea. The phrase is not the important part. The useful part is the observation that distributed action can emerge when many actors receive the same class of inputs and each actor has an expected role. The system does not need one mastermind at the center of every move. It needs a stable environment, legible signals, and clear boundaries.

Why Human-Centred Still Matters

This does not mean the human becomes irrelevant. It means the human moves up a level.

The strongest practical position for the near term is not “agents should run everything” and not “humans should manually do everything.” It is this:

Human-framed. AI-expanded. Human-authorized.

That principle is strongest when applied to new features, new products, new workflows, and new organizational moves.

In software, for example, AI can help a lot. It can generate code, propose edge cases, surface alternatives, and compress the cost of exploration. But if a feature is important enough to alter customer experience, cost structure, support load, or product direction, then the feature should still originate in a human frame. A human should be able to explain the reason for it, the tradeoff behind it, and the responsibility for shipping it.

That is not a romantic position. It is an operating discipline.

There are good reasons for it:

  • humans still carry legitimacy
  • humans still absorb accountability
  • humans still hold tacit context about exceptions and politics
  • humans still learn by framing problems, not only by approving outputs
  • humans still need to understand why something exists if they are expected to own it later

The human-centred principle also works as a continuum, not a binary. Reworking Order should not pretend that all meaningful change must always be manually initiated forever. A better model is a ladder:

Human-initiated, human-authorized
Human-initiated, AI-assisted, human-authorized
AI-suggested, human-initiated, human-authorized
AI-suggested, human-reviewed, human-authorized
Agent-initiated within a bounded policy, human-reviewed
Agent-initiated within a bounded policy, exception-raised to a human

The practical question is not whether the organization believes in humans or agents. The practical question is where on that ladder a given type of work belongs.

For high-consequence work, the answer should usually stay near the top. For low-risk, repetitive, or tightly bounded work, the answer may move lower over time. That is not a philosophical compromise. It is a design choice based on risk, confidence, and accountability.

Reading The Traces

If work is distributed, then the most important evidence is not a single order. It is the trace.

Traces are the artifacts left behind by action:

  • prompts
  • edits
  • comments
  • logs
  • approvals
  • redlines
  • version history
  • tickets
  • handoffs
  • notifications

If you want to understand what the system actually is doing, you do not look only at the stated policy. You look at the traces. This is where a few older theoretical ideas become useful without becoming the whole show.

Foucault is relevant here because power is not only what someone says out loud. It is also how a system arranges visibility, permission, and discipline. Actor-network theory is relevant because agency is not concentrated only in the human manager; it is distributed across people, software, documents, incentives, and rules. Baudrillard is relevant as a warning: once dashboards, summaries, and synthesized outputs become the thing people trust, there is a risk of mistaking representation for reality.

For business use, the point is simple. A modern operating model should be traceable. If an agent did something useful, you should be able to reconstruct:

  • what signal triggered it
  • what policy allowed it
  • what human was responsible
  • what evidence it used
  • what got reviewed
  • what outcome it produced

That is how distributed work becomes governable. It is also how trust is built.

In other words, traces are not just forensic. They are managerial.

What This Means For Reworking Order

Reworking Order should not present AI work as a story of central control. It should present it as a story of designed emergence.

The practice thesis is not that agents should replace people. It is that work should be reorganized so the right parts of the system are human-led, machine-augmented, and policy-bounded. That creates room for speed without losing authorship.

That is also where the “standalone complex” intuition becomes practically useful. In some environments, a group of actors with similar inputs and expected roles will behave coherently without one central commander. That is true in markets, in teams, in distributed operations, and increasingly in human-agent systems. But coherence only happens when the environment is built well:

  • roles are clear
  • signals are legible
  • exception paths exist
  • accountability is visible
  • the system leaves traces

If those conditions are missing, distributed systems do not become elegant. They become noisy.

So the operating question for the business is not, “How do we put agents at the center?” The better question is:

What environment, signals, and trace structure produce coordinated work while preserving human judgment and responsibility?

That is a more precise thesis than a simple boss-subordinate model. It is also closer to what intelligent buyers actually need. They do not need a theory of robot autonomy. They need a way to redesign work so that distributed action becomes reliable, legible, and accountable.

That is what a standalone complex of work really means in practice: not a single commander, but a coherent field of actors responding to signals, carrying roles, and leaving traces that make the whole system intelligible.

Diagram

INTENThuman intentSIGNALSshared signalsFIELDenvironment / policyinterfaceAIAI suggestionsROLElocal role behaviorTRACEtracesREVIEWhuman review / learningReworking Order
Coordination emerges from shared signals, environment, and traces rather than a single commander. Humans and agents both act locally inside a designed field.

Sources

Supporting research and framing:

  • https://www.nber.org/papers/w26673
  • https://www.nber.org/papers/w33949
  • https://academic.oup.com/pnasnexus/article/5/3/pgag030/8490283
  • https://www.nature.com/articles/s41562-024-02024-1
  • https://www.nber.org/papers/w34859
  • https://www.nber.org/papers/w33795
  • https://www.nber.org/papers/w28811
  • https://www.nber.org/papers/w25148
  • https://www.nature.com/articles/s41598-026-42312-6
  • https://www.nber.org/papers/w31325

Conceptual lenses mentioned in the essay:

  • actor-network theory
  • Foucauldian power as arrangement rather than just command
  • Baudrillard on representation and simulacra as a caution against mistaking dashboards and summaries for reality