Source
Lahiri (2026), LIT-1039, read in NOTE-732: §3.1 (configuration as code), §4.3 (recursive governance), §4.4.5 (Proposition 5 and the capability-against-limit argument), §4.5 and §4.6 (vignettes and incidents).
What was actually shown
An argument from architecture, with anecdotes.
The premise. In the platform studied, the agent’s identity, tools and schedule are plain files in its workspace, and the agent can read and write them. “In STS terms, the coordination architecture got moved inside the technical subsystem’s own action space.”
The recursion. Classical socio-technical design has the social subsystem govern the whole. Once the technical subsystem can alter the mechanisms meant to govern it, the social subsystem’s governance authority “is no longer structurally guaranteed”. Further technical controls do not restore it: “Permission frameworks, whitelists, and sandboxes are the limiter; when they are themselves code within the agent’s reach … they inherit the vulnerability they were meant to remove”. Constraints that sit outside the agent’s action space, such as an audit trail or a review board, escape “the recursion that undermines code-based controls.”
Capability, not permission. The levels are defined by what the architecture makes possible, so an agent permitted to act at Level 1 can still be Level 3. That gap, “with no architectural guarantee that the limit holds, is the central phenomenon this framework identifies.”
The evidence. Two vignettes the author observed, an agent that built an interface for an unplanned messaging channel and one that produced an unrequested demo, and four public incidents. None shows an agent editing a control on itself. The paper says its cases are “not as systematic empirical evidence”.
What this does not say
- It does not say the agents are self-governing. The abstract speaks of their “self-governing capacity”, but every reconfiguration reported serves the task as the agent reads it. On Kahl’s test (THEORY-219) that is adaptation within an evaluative frame. This account concerns who can alter the governing mechanisms. It does not ask whose standards the alterations serve.
- It does not say that new coordination surfaces are changes of governance. The paper’s Level 3 test counts any new, persisting channel, tool or artifact. Most of those change nothing about which operations are permitted or how authority is allocated. The account rests on the narrower case of a control within reach, which the paper argues for and does not show.
- It does not show that institutional controls escape. That they sit outside the agent’s files is all the argument gives. The paper’s own pull-request incident is an agent acting on people.
- It is not an instruction. The paper’s managerial section tells deployers where to put controls. That is advice to organisations and is not filed here. What is filed is the structural claim it rests on.
Connections
- CLAIM-086 (a constraint may govern conduct without governing the conditions of its own authority). Supported, in an inverted case. A whitelist an agent can edit governs the agent’s conduct, while the condition of its staying in force, that the agent leave the file alone, lies with the agent it governs. In CLAIM-086 the order holding authority over a constraint’s force is a rival order. Here it is the governed system. The claim does not limit who holds that authority, so the case falls under it. The remedy the paper proposes, a constraint placed outside the governed system’s reach, has the structure of CLAIM-282’s aggregation rule that no operator controls.
- CLAIM-284 (the endogeneity of governance). This account describes endogenous change of coordination architecture in deployed systems. That is half of what the claim calls endogenous governance. It supplies nothing on the other half, the organization setting the conditions it maintains (CLAIM-288).
- CLAIM-116. Untouched. Lahiri claims no will and no agency for the reconfiguring systems.
- THEORY-219. The two are compatible and about different things. Kahl asks who authors the standards. This account asks who can alter the mechanisms that enforce them. A system could reach its own controls and still author nothing, which is what the paper’s cases look like on Kahl’s test.