This preprint starts from an asymmetry: coding agents let digital execution — repository analysis, code generation, testing, migration, tool use and parts of operations — be replicated and parallelized without a matching growth in headcount, while problem framing, semantic commitment, verification, integration, human attention and the acceptance of residual risk remain bounded by human cognition, organizational authority and budget. The authors argue that this moves software engineering’s object of control away from human activity and toward the software change itself, a shift they summarize as moving from human process control to trusted change governance.
To organize that shift the paper proposes two primary constructs and one execution abstraction.
Trustworthy Change is the engineering object that follows a change from intent through
delegation, verification, integration, acceptance and operation. Responsibility Topology
classifies a software organization by how independent authority to accept residual risk is distributed,
distinguishing a single-center form with one final responsibility anchor from a multi-anchor form that
requires joint acceptance across independently governed domains. The
Human–Agent Cell is the execution unit — a human with their agents, context, tools,
permissions and budget — which produces candidates, proposals and evidence but does not by executing
gain authority to accept them. A running example, splitting an order system’s cancelled status into
user cancellation and payment-timeout expired, is used throughout to show that a small artifact change
can carry a large semantic, verification and responsibility scope.
The authors describe their contribution as theory construction and operationalization rather than empirical validation, and state explicitly that responsibility, accountability, change management, specification, verification and human oversight all predate the paper; their narrower claim is that agent-scaled execution changes how these concerns fit together.
Key Points
- The paper represents a change’s governance as five loops — intent, delegation, verification, integration and operation — with specification as the refinement that makes intent delegable, evidence as the main output of verification, and responsibility cutting across all five.
- It separates a change’s progression into Candidate, Eligible and Accepted states, so that finishing implementation, passing engineering qualification and formally accepting residual risk are not collapsed into one notion of “done”; only an accepted change may modify the authoritative baseline.
- It lists six conditions that stop a candidate from becoming eligible — failures of intent/specification, delegation/resources, change/provenance, evidence, integration/operation and responsibility — and treats a change whose required approver exists but has not yet signed off as acceptance-pending rather than as a responsibility failure.
- It argues that Responsibility Topology is independent of headcount: a founder running twenty concurrent agents can remain single-center, while two engineers can form a multi-anchor organization if each holds a non-revocable acceptance right over an affected domain.
- It distinguishes pressures created by distributed execution (context divergence, stale context) from those created by distributed authority (joint acceptance, responsibility closure), and proposes Context Invalidation — each in-flight task choosing to continue, mark stale, refresh, replan, reverify or stop when an authoritative fact changes — as the response to the first.
- It proposes a Minimum Sufficient Specification hypothesis: that for a given class of change there may be a level of specification rigor beyond which added rigor no longer repays its lifecycle cost; the cost curves it draws are labelled as illustrative and not fitted to data.
- It treats a task given to an agent as a contract covering goal, context, scope, permission, budget, acceptance, stop condition and handoff, and makes the stop condition first-class so that an agent escalates rather than proceeding when it needs unauthorized data, wider write access, an irreversible migration or a larger budget.
- It frames agent security around containment: a compromise of an agent’s interpretation, for example by prompt injection, should not become an unbounded compromise of its actions, which permission boundaries, sandboxes, tool hooks and egress policies should limit.
- It argues that when generation outpaces verification a verification queue forms, and that as it grows the responsible person is pushed toward relying on summaries and rapid approvals; it proposes counting accepted changes, rather than lines of code, commits or agent tasks, as the accounting boundary for trustworthy output.
- It states eight propositions with observable expectations and falsification signals — among them that agent concurrency saturates against fixed verification capacity and that Responsibility Topology keeps explanatory power after controlling for headcount — and proposes an “augmented Conway” hypothesis that agent delegation and context dependency add explanatory power for software co-change beyond human communication.
Notes
The paper positions itself against SE 3.0 and Structured Agentic Software Engineering (SASE), which it describes as examining changes to actors, processes, tools and artifacts, and against Spec-driven development and harness engineering, whose primitives — repository instruction files, skills, hooks, MCP tools, sandboxes and sub-agents — it maps onto its own constructs rather than claiming as new. It also describes adjacent 2026 work on human–agent responsibility protocols and graduated oversight as specifying how responsibility is constrained or oversight calibrated, rather than classifying organizations by independent acceptance centers.
The authors note that their running example is illustrative and not a case study, that their capacity bounds are schematic rather than calibrated laws, that the empirical commitments cover only the single-center and multi-anchor forms and not hierarchical, federated or platform-mediated responsibility, and that any of their central distinctions may fail to survive measurement.