AI-Assisted Software Engineering as Hostage Negotiation

Supervising an autonomous coding agent can feel less like programming and more like a hostage negotiation. Something you value, your codebase, is in the hands of a party that acts on it directly, while you can only influence how it acts. This paper takes that feeling seriously, as a structural allegory and not a claim about machine intent: the agent is not malicious and wants nothing. What the allegory captures is the shape of the work. A hostage negotiator does not command the other side; the craft is to build a working relationship with an unpredictable party and, through it, bring the hostage out intact. The engineer’s task has the same shape. Set terms the agent will hold, read what it does against what it says, calibrate trust turn by turn, and steer a volatile collaboration toward a result you can live with. Naming that shape explains a frustration that prompt engineering and context engineering leave untouched, and it points to a skill the field has not yet named, a relational discipline and a demanding one: getting good work out of a capable, unreliable collaborator that has nothing to lose and whose hands are on something you cannot replace.