When the environment starts coordinating the agents, how do we govern the actions?
An agent leaves a research note. Another turns it into a task. A third changes software. A failed test prompts a repair, and a release process carries the result into the world.
Each step may look reasonable on its own. Together, they can produce activity that no single participant planned or fully understands.
Where, in that arrangement, is the brain—and where should human control sit?
From the termitary to stigmergy
In The Soul of the White Ant, Eugène Marais approached the termite colony as a composite animal. He also attributed a directing psychological power to the queen. That historical explanation should not be mistaken for an established mechanism or imported into computing. His enduring question is how coordinated activity becomes intelligible when we look beyond individual workers. [1]
Stigmergy is coordination through traces left in a shared environment. An action changes the surroundings; that change influences what the same participant or another does next. Francis Heylighen describes this relationship between action, agent, and medium. It is an established concept with applications beyond biology. [2]
Termite construction offers a concrete example. Facchini and colleagues studied environmental cues in building, finding support for the role of surface geometry and proposing evaporation as the mechanism through which it is sensed. This is evidence about particular construction conditions, not a complete account of colony behaviour. [3]
The useful parallel is that a structure can help organize its own further construction.
For software, agentic stigmergy describes agents responding to persistent traces—documents, task records, code changes, or other shared state. A shared folder alone is insufficient: what is left there must influence subsequent action.
The concern: environmental intelligence
We use environmental intelligence here as a working description of capability distributed across agents, persistent records, triggers, and the rules connecting them. It does not mean that the environment is conscious, has intentions, or constitutes a newly established scientific category.
The concern is that the environment becomes an active part of how work is selected and continued.
A hypothetical product team illustrates the difference. A research agent records a customer complaint. A planning agent treats it as a priority. A developer changes the product. A release agent acts on a successful test. Nobody explicitly approved that entire chain; each participant responded to the state it encountered.
A centrally controlled workflow could produce the same sequence. Direct messages could coordinate it too. The stigmergic element is the use of persistent environmental changes as cues for further action. Real systems may combine these mechanisms.
That persistence can be valuable: work survives handoffs, agents discover unfinished tasks, and useful knowledge accumulates. But an unsupported assumption can also become a task, then a change, then an apparent precedent. Repetition across agents does not supply independent evidence.
The environment can preserve both useful knowledge and reasons to keep making the same mistake.
What the evidence does—and does not—show
The SwarmWorld preprint studies language-model agents in a simulated shared environment. Its authors report broader and more resilient portfolios of constructed technologies than a matched isolated-search comparison, while isolated search remains competitive for the strongest individual artifact. This is bounded evidence for accumulated shared construction, not proof that agent collectives are universally better or beyond control. [4]
Research by Cemri and colleagues also documents failures in multi-agent system design, alignment between agents, and verification. [5]
Our governance concern is an inference to investigate: if shared traces help organize activity, understanding individual agents may leave important causes of that activity unseen.
Tracking agents is a starting point
Knowing which agents are active is useful. Governing their effects requires following the connection between an action, its authority, and the state that prompted it.
Consider an agent that changes a customer record. An activity log may identify the writer. We still need to know which source justified the change, which rule permitted it, whose authority applied, and which downstream processes it activated.
A proposed governance tool should help answer five questions:
- Origin: What request, evidence, or environmental change prompted this action?
- Authority: Which identity, permissions, and applicable policy allowed it?
- Consequence: What changed, and what further activity could that change trigger?
- Intervention: Where can the action be denied, paused, or escalated before its effects occur?
- Continuity: Can the same work resume through another agent, retry, or persistent task after cancellation?
These are proposed requirements, not claims about an existing product. Logging after execution cannot by itself enforce authorization. For governed actions, policy checks must operate at the points where agents can change systems, spend resources, or delegate work. Any guarantee would be limited to the actions and execution routes those controls actually cover.
Stopping the agent may leave the work alive
Imagine that an agent starts an unwanted sequence. We stop it, but another agent finds its unfinished tasks. A scheduler retries the job. A separate service reacts to an output already delivered.
This hypothetical scenario needs no desire for survival. Ordinary persistence and retry rules could sustain the process.
In a bounded environment, operators may be able to revoke permissions, disable triggers, and quarantine the records that would restart activity. Across independent systems, authority may be fragmented. Some consequences may already be irreversible.
The test is therefore more demanding than whether one agent stops responding: can we identify the state sustaining the activity, block its continuation within the systems we control, and verify that a restart does not revive cancelled work?
Visibility, enforceable authority, and effective intervention need to be examined separately.
Why this matters to HumainX
The Neo-Cottage Revolution asks whether AI can reduce the minimum economically viable size of many knowledge-based firms. Environmental intelligence suggests a refinement: fewer human employees may coexist with a substantial organization of software, records, dependencies, and controls.
Whoever controls that environment may control important parts of production. A small firm might gain capability while depending on someone else's memory, permissions, and infrastructure. Portable records and replaceable providers could support a different outcome.
The human economic role may include deciding which outcomes are legitimate, granting authority, and accepting responsibility for the collective result. Whether those functions remain scarce—and who will pay for them—remains open.
A practical next investigation
We plan to build a tool to track agents and help ensure their actions are governed. HumainX will use that work as a subject for inquiry: what can such a tool actually observe, enforce, and stop? Building and applying the technology belongs to Mindsgate's operational work; HumainX's task is to examine its implications and limitations in public.
A useful experiment would cancel a task, replace the responsible agent, and test whether shared records or retries cause the activity to resume. It should record both successful interventions and gaps in coverage. A tracking dashboard is not evidence of control until interventions have been tested.
The concern is not that we have demonstrated an unstoppable collective. It is that our unit of oversight may be too small.
If intelligence increasingly resides in the relationship between agents and their environment, can we govern the whole chain of action—or only watch its individual parts?
Sources and scope
This is an exploratory argument. The business examples and shutdown scenario are hypothetical; the economic propositions are hypotheses. No claim of collective consciousness or inevitable loss of control is made.
- Eugène Marais, The Soul of the White Ant, especially Chapter 12, “The Mysterious Power which Governs”. Historical primary text; chapter references avoid edition-dependent pagination.
- Francis Heylighen (2016), “Stigmergy as a universal coordination mechanism I: Definition and components”, Cognitive Systems Research, 38, 4–13.
- Giulio Facchini and colleagues (2024), “Substrate evaporation drives collective construction in termites”, eLife. Supports the bounded construction example, not the software or economic conclusions.
- Subhadeep Pal, Fiona Y. Wang, and Markus J. Buehler (2026), “SwarmWorld: Stigmergic technological evolution in societies of language-model agents”. Preprint; simulated systems, with outcome-dependent advantages.
- Mert Cemri and colleagues (2025), “Why Do Multi-Agent LLM Systems Fail?”. Evidence of failure mechanisms in the systems studied; not a forecast of all future deployments.