Who is in charge? Coordination and governance in agentic software engineering

AI agents are transforming software development from a human-led activity into a collaborative process involving autonomous planners, coders, testers, and reviewers. This shift raises a deceptively simple question with deep engineering and social consequences: when software work is shared between humans and autonomous agents, who is really in charge? This paper argues that the central challenge is coordination: deciding who does what, when, and under whose authority. Natural-language conversation is now the primary medium through which humans and agents interact, assign tasks, and negotiate outcomes; however, conversation is not a substitute for explicit coordination structure.
We trace the evolution from Agent-Oriented Software Engineering (AOSE) to contemporary Agentic Software Engineering (ASE), and use a multi-agent code-review scenario to distinguish conversational interaction from governed coordination. The paper argues that autonomy increases the need for coordination; it does not reduce it. Making agentic systems governable therefore requires explicit mechanisms for assigning responsibility, validating outputs, escalating uncertainty, and preserving accountability.
We introduce the notion of a runtime coordination contract, a lightweight governance abstraction that specifies roles, task scopes, input/output schemas, escalation conditions, and audit obligations. The paper positions coordination contracts as a runtime analogue of classical coordination and behavioral-contract mechanisms: more flexible than fully pre-designed protocols, but sufficiently explicit to support monitoring and auditability. We discuss some implications for software engineering practice. We conclude proposing an open research agenda concerning contract languages, runtime conformance monitoring, multi-agent benchmarks, and human–agent teaming.

keywords   Multi-agent systems, Coordination, Large language models, Governance, Conversational agents

journal or series

 Journal of Logical and Algebraic Methods in Programming (JLAMP)