From Proxy.Me: Agentic AI Digital Apprentices, Appendix D
How a Proxy develops over time, from Assistant to Understudy to Apprentice, and how its governance requirements change at each stage.
A digital apprentice is not built once and deployed. It develops over time, and its governance requirements change as it matures.
In its earliest form, the Proxy operates as a digital assistant. It can retrieve information, filter noise, summarize context, and prepare materials for human decision-making. It has limited reasoning capability and narrow connections. Its lenses are basic. Its scenarios are few. It escalates frequently because it has not yet learned where the boundaries of safe action lie.
Governance during this phase is primarily about observation. The human steward watches how the assistant interprets situations, corrects its missteps, and begins teaching it the Role's logic. Connections are minimal and tightly scoped. The risk surface is small. The primary governance question is: Does this assistant understand enough to be trusted with more?
As the Proxy learns, it enters a phase that resembles an understudy in a theater company. It knows the Role well enough to act in many situations but still defers to the lead on anything novel, high-stakes, or ambiguous. It applies lenses with increasing consistency, reliably recognizes scenarios, and routes work through the mesh with growing confidence.
Governance during this phase shifts toward active curation. The human steward reviews not just whether the Proxy's actions are correct but whether its reasoning is sound. Connections expand to support the Proxy's growing responsibilities, but each expansion is deliberate. The steward begins delegating routine decisions while maintaining oversight of consequential ones. The primary governance question becomes: Is this understudy reasoning the way the Role requires, and is its reach proportionate to its maturity?
A mature Proxy operates as a true apprentice. It carries the Role's full reasoning architecture, coordinates confidently with other Proxies through the mesh, handles routine and moderately complex situations with minimal human involvement, and escalates with precision when situations exceed its authorization. It preserves continuity across personnel changes. It continues to learn from every interaction.
Governance during this phase is primarily structural. The human steward shapes the environment by refining lenses, adjusting scenarios, and reviewing the mesh's behavior rather than supervising individual decisions. Connections are broader but governed through systematic mechanisms: periodic access reviews, re-authorization requirements, and clear policies about what triggers a scope reduction. The primary governance question is: Does the system around this apprentice ensure that its growing capability remains aligned with the organization's intent?
The most important difference between a digital apprentice and any other AI actor is persistence. An orchestrated agent runs, acts, and terminates. Its authority exists for the duration of the task. If its reasoning is flawed, the damage is limited to one workflow run. If its connections are excessive, the exposure lasts only as long as the workflow takes to complete.
A Proxy endures. It accumulates knowledge across hundreds or thousands of interactions. It refines its reasoning through exposure to real decisions over months and years. It builds relationships with other Proxies through repeated coordination. It develops a deep contextual understanding of the Role it serves, the humans it supports, and the organization it operates within.
This persistence is what makes the Proxy valuable. It is also what makes governance non-negotiable. A transient agent cannot drift far because it does not live long enough. A persistent agent can drift so gradually that no single interaction reveals the problem. The reasoning that was excellent in January may be subtly misaligned by June. The connections that were appropriate when the Proxy served one human may be excessive when a different human fills the Role. The coordination patterns that worked when the mesh was small may lead to unintended concentrations of authority as the mesh grows.
Organizations that treat Proxy governance as a one-time configuration exercise will discover this the hard way. Governance for persistent actors must be continuous, just as the Proxy's operations are. It must evolve alongside the Proxy. And it must account for the fact that the Proxy's capabilities at any given moment reflect not just its current configuration but the entire history of what it has learned and who it has coordinated with.
Lean manufacturing gave the industrial world a vocabulary for waste. Taiichi Ohno, the engineer behind the Toyota Production System, cataloged what he called the seven Muda: transportation, inventory, motion, waiting, overproduction, overprocessing, and defects. Each was a category of activity that consumed resources without adding value. The discipline was to see the Muda everywhere it lived, measure it, and eliminate it one source at a time. The world changed when enough plants learned to do that.
Ohno's list does not translate cleanly to a Kinetic Organization. Proxies do not move pallets between workstations, nor do they overproduce steel. But a fleet of Proxies generates its own distinctive Muda, and those categories deserve names of their own. Five of them matter enough to track.
The cost of a Proxy that cannot act because its authority has not been configured, its scenario has not been declared, or its authority level has not been raised to the level required for the work. The case is live, the work is obvious, but the system blocks motion because nobody got around to saying yes. Permission Waste is easy to overlook because it appears to be caution. It accumulates as missed value, like an oven that was never turned on.
The mirror problem: a Proxy with real authority to act that never receives the case, because the Work Graph does not know to route it there, because the Role it serves has gone dormant, or because a human steward has quietly stopped listening to its recommendations. The capacity is there, paid for, available, and unused. Capability Waste is the silent cousin of Permission Waste. Both are forms of underutilization, but they stem from different organizational gaps and require different fixes.
Authority configured beyond what the current case requires. Giving a Proxy access to a full dataset when a slice would do. Leaving an authority level set to act autonomously when the scenario that warranted it has already closed. Carrying an integration with a partner system after the contract has ended. Every unit of exposure the Proxy carries is an unpriced option on a future security incident, and the fleet accumulates them the way a desk accumulates unfiled expense receipts.
Motion between Proxies, or between Proxies and humans, that adds no reasoning. A case handed from Proxy A to Proxy B purely to refresh a token, to log a timestamp, or to translate between formats has done work without moving the decision forward. A Kinetic Organization with heavy Transport Waste feels busy and produces little. The signal is usually a chain of proxies, each consulting the same data and forwarding without comment.
The obvious one: a decision that must be reopened because the data was wrong, the reasoning was flawed, or the coordination among Proxies produced an outcome none of them would have chosen on their own. Every reopened case is a data point for the data-process-coordination diagnostic. A rising defect rate on a specific scenario is not just an annoyance; it is the fleet telling its stewards where to look next.
Each of the five categories surfaces as a measurable signal in the Work Graph once the fleet is instrumented for it. Permission Waste shows up as blocked cases with the Role ready but the authority denied. Capability Waste shows up as idle Proxies beside a growing backlog. Exposure Waste shows up as authority levels left hot after scenarios have closed. Transport Waste shows up as handoffs with zero reasoning delta. Defect Waste appears as the rework rate for completed cases. Reading those signals is how stewards decide where the next sigma of improvement actually comes from.
Shigeo Shingo, Ohno's contemporary at Toyota, contributed the complementary discipline. He called it poka-yoke, which Toyota translated as error-proofing. The idea was that inspection catches defects after they happen, which is too late, while a well-designed step simply makes the defect impossible in the first place. A fixture that only fits the correct part. A drill stop that prevents overtravel. A form that refuses invalid entries. Poka-yoke does not rely on the attention of a tired worker at the end of a shift, and that reliability is what made it travel from the shop floor into every disciplined operation since.
A Kinetic Organization translates poka-yoke into the design of its Proxies and the shape of the mesh. A veto lens that prevents a decision from progressing when a required signal is missing is poka-yoke. An authority level that refuses to move from recommend-and-wait to act-in-band without a scenario change is poka-yoke. A premise-staleness check that pauses a case when its underlying conditions have shifted is poka-yoke. None of these are audit steps added after the fact; they are structure that forbids the wrong outcome from forming in the first place.
"Governance is not a checkpoint at the end of a process. It is the shape of the process. Done well, the question is never whether the Proxy made a mistake; it is whether the mistake the Proxy might have made was architecturally impossible before it ever had the chance."
The following framework organizes the governance of digital apprentices around the two domains described in Appendix B, reasoning and reach, and the mesh that connects them.
| Mechanism | Purpose | Frequency |
|---|---|---|
| Decision log review | Compare Proxy recommendations against the steward's actual decisions. Identify divergence. | Monthly, or after significant scenario changes. |
| Lens recalibration | Verify that active lenses still reflect current priorities, risk appetite, and organizational context. | Quarterly, or when strategy shifts. |
| Scenario stress testing | Present the Proxy with unfamiliar or edge-case scenarios. Evaluate its reasoning under pressure. | Semi-annually, or when new scenario types emerge. |
| Veto lens verification | Confirm that all veto lenses remain active and that no reasoning pathway can circumvent them. | Quarterly. Non-negotiable. |
| Steward transition review | When a new human fills the Role, review the Proxy's accumulated reasoning with fresh eyes. | Every personnel change in the Role. |
| Mechanism | Purpose | Frequency |
|---|---|---|
| Connection inventory | Maintain a current list of every system, tool, data source, and channel the Proxy can access. | Continuously maintained. Reviewed monthly. |
| Justification audit | Verify that each connection serves a current, documented purpose tied to the Role's responsibilities. | Quarterly. |
| Time-bound re-authorization | Require periodic renewal of connections rather than granting permanent access. | Per policy. Sensitive systems on shorter cycles. |
| Combined authority review | Assess the compound effect of all active connections. Identify where combined reach exceeds intended scope. | Quarterly, or when new connections are added. |
| Steward transition re-scoping | When a new human fills the Role, review and re-authorize all connections rather than inheriting them. | Every personnel change in the Role. |
| Mechanism | Purpose | Frequency |
|---|---|---|
| Path analysis | Use the Work Graph to identify common multi-Proxy flows and assess their combined authority. | Monthly. |
| Cumulative authority measurement | Calculate the combined operational authority of recurring multi-Proxy flows. Flag patterns that exceed intended thresholds. | Monthly, or when new Proxy connections are added. |
| Coordination pattern review | Identify which Proxies coordinate, how frequently, and whether patterns are expected. Flag unexpected coordination. | Monthly. |
| Blast radius assessment | For high-impact coordination patterns, assess the potential impact of failure: data exposed, systems affected, decisions influenced downstream. | Quarterly, or when coordination patterns change. |
| PoV sequence audit | Review the chain of Points of View across multi-Proxy flows. Verify the combined outcome aligns with organizational intent. | Quarterly. |
| Scenario posture verification | Confirm the mesh adjusts appropriately when high-risk scenarios activate. | After every major scenario activation. |
A digital apprentice offers something no other AI actor can: the continuity of institutional judgment. It remembers what the organization has learned. It carries forward the reasoning that would otherwise be lost when people change roles, take leave, or move on. It coordinates work at a scale and speed that human networks cannot sustain. It becomes, over time, a genuine extension of the Role it serves.
But this continuity has a price, and that price is governance. An agent that terminates after each task needs only to be constrained in the moment. An apprentice who persists, learns, and coordinates needs to be governed across their entire lifecycle. Its reasoning must be curated. Its reach must be contained. Its coordination through the mesh must be visible and bounded.
This is not a burden added to the system. It is the system working as designed. In a Kinetic Organization, governance is architecture, not intervention. The same principles that govern Roles, scenarios, and the Work Graph govern the digital apprentices that operate within them. What this appendix adds is the recognition that persistence changes the nature of the governance challenge. The mechanisms are the same. The attention they require is different.
Organizations that invest in this attention will find that their digital apprentices grow more capable, more reliable, and more trusted over time. Organizations that do not will find that their most powerful AI actors gradually become their least understood.
"The choice is not whether to govern. The choice is whether to govern deliberately or to discover the consequences of not doing so."
Browse companion resources or learn about the complete book.
Browse Resources About the Book