What it is
Agent handoff is the term for the moment an autonomous AI system passes a task to a human instead of completing it alone. The pattern shows up wherever autonomous systems operate: a support bot escalates to a live rep when a query falls outside its script, a multi-agent framework passes a thread from one agent to another with full context attached. In every version, the system reaches a boundary it was told not to cross, and control passes to whoever set that boundary.
Applied to an AI sales employee running hiring-signal detection, scoring and outreach drafting, an agent handoff is the specific instant it stops short of sending, scoring or acting on its own, and instead surfaces the task to the consultant with context attached, because the situation sits outside what it is authorised or confident enough to decide alone: an unfamiliar account, a message that touches an existing client relationship, a score sitting right on the ICP line, or a request that brushes up against commercial terms.
An agent that never hands off is not autonomous. It has just not hit its boundary yet.
Why it matters
Autonomous AI only becomes safe to run unattended once it knows precisely when to stop, not just what to do. A system that never hands off either acts past its brief the moment reality gets unusual, or gets built so cautiously that everything gets flagged, and the human ends up doing the reviewing work the automation was meant to remove. The handoff is what lets both failure modes cancel out: the agent runs wide autonomously, and narrows to a person exactly where judgement, not more processing, is what is actually needed.
It is also what makes an "autonomous mode" claim credible rather than marketing. The right question for an agency evaluating an AI sales tool is not "can it run without me" but "what exactly makes it stop and hand back to me", because that boundary is the difference between a system that is genuinely autonomous and one that is simply unsupervised and has not made a visible mistake yet.
How boilr handles it
In boilr, an agent handoff is triggered by the same signals that already govern autonomous mode: an AI confidence score dropping below the threshold you have set, a guardrail boundary the task would otherwise cross, or a pattern the Company Brain has not seen enough of yet to act on with confidence. When one of those fires, your AI sales employee does not guess and does not block the rest of the pipeline. It hands that one task into your Tasks inbox, flagged with the reason it was escalated, while everything else within its authority keeps running.
Every handoff you resolve, approve, edit or reject, feeds back into the Company Brain, so the boundary itself gets sharper over time: fewer handoffs on the patterns you consistently approve as drafted, and the same caution held wherever your desk actually needs it. The handoff is not a limitation you tolerate. It is the specific feature that lets you trust the rest of the system to run without you watching it.
