The Human–Agent Cell (HAC) is the execution unit proposed in Software Engineering in the Agent Era: From Trustworthy Change to Human–Agent Software Organizations because, in that paper’s words, the word “developer” no longer fully describes the execution unit that actually performs software work once agents are involved. A HAC centres on one human execution subject and combines that person’s set of personal agents, their local working memory or context, the available toolset, delegated permissions and a resource budget. It may use one agent or run several specialized agents concurrently; changing a model version, adding agents or changing a token budget does not usually create a new HAC, whereas a material change in the human subject or the formal engineering boundary is treated as a handoff, termination or redefinition of the cell.
Usage
The paper gives the HAC a minimal formal interface. Its inputs are a snapshot of the authoritative engineering state, a task-and-resource contract and signals that the authoritative state has changed; its outputs are a candidate change, proposals to change the authoritative state, and local evidence. Inside that boundary the cell may hold substantial local information — scratch memory, provisional analyses, debugging hypotheses, unaccepted designs — but this is execution state, not organizational truth: a proposal becomes an organizational fact only through authorized acceptance followed by a versioned commit.
The central point of the abstraction is the separation between execution and acceptance. A HAC produces work, but execution alone does not confer authority to accept it. That authority belongs to responsibility anchors under the organization’s Responsibility Topology; a HAC holds final acceptance authority only if its human subject also holds the required formal authority. The same person can be both a HAC’s execution subject and an anchor for some domain, or can execute under another anchor’s authority, and the paper keeps the two counts distinct. Several HACs working in parallel will naturally hold different contexts, so their candidates can each be locally coherent yet mutually incompatible — the pressure the paper answers with shared authoritative state and context invalidation.
When It Applies
The paper studies the HAC as the human-centred execution boundary in agentic software work, in which one human execution subject works with one or more agents. It does not claim it is the only possible execution unit: team-level, organization-level or fully autonomous agents may exist outside any HAC while remaining subject to the same distinction between execution authority and acceptance authority. It also cautions that separate HACs are an execution boundary, not a guarantee of independent judgment, since different cells may share models, tools and retrieval sources and so fail in the same way. Like the paper’s other constructs, it is proposed as part of a theory awaiting empirical testing.