Agent Handoff — Stack · Agent Architecture and Orchestration

Technology landscape as of 2026-10-10. Knowledge cutoff May 2026.

Selection criteria

Almost every criterion here is about the payload: what it can hold, how faithfully, and what it costs to carry.

Payload representation. Whether the transferred package has a declared schema that can be validated and tested, or is whatever the sender decided to write this time.
History fidelity against context cost. Carrying the whole transcript is faithful and expensive; carrying a summary is cheap and lossy. The mechanism should let you choose per field rather than globally.
Evidence access. Whether the receiver can reach the sender's tool results and retrieved documents, or only the prose that was derived from them.
Permission transfer. Whether tool scopes and data entitlements are explicit in the transfer or inherited implicitly from the runtime.
Return path. Whether the mechanism distinguishes delegation that comes back from custody that does not.
Process boundaries. Whether a transfer can cross processes, services, or organisations, and what identity and durability that requires.
Cycle prevention. Whether hop counts and visited-agent sets are carried and enforced by the mechanism, or left to each agent's instructions.
User-visible semantics. Whether the surface can represent a transfer at all, or whether everything must look like one continuous speaker.

Option	Best for	Watch out for
Framework transfer primitives (agent-SDK handoffs, graph-runtime goto and command constructs, speaker transitions)	Single-process agent systems where all participants share a runtime	What is carried is usually the whole message list by default — faithful, expensive, and unfiltered
Sub-agent invocation that returns control	Delegation where the original agent remains accountable for the user	Not a custody transfer; the caller keeps paying context cost for the whole exchange
Transfer by reference to shared state	Long-running work where the payload is too large or too sensitive to copy	Requires a durable store and a custody rule, or two agents will write the same state
Message queue with a typed transfer envelope	Agents owned by different teams, deployed independently	Asynchrony changes the user-facing experience; the sender must say something while the receiver starts
Cross-organisation task protocols (A2A and similar)	Transfers that leave your trust boundary entirely	Identity, authorisation, and data-residency questions precede the technical design
Contact-centre and ticketing escalation paths	Transfers whose destination is a human queue with its own tooling	The payload format is dictated by that system and is usually a text note field
Model-written summary passed as the payload	Prototypes and low-stakes transfers	Non-deterministic content, untestable, and the first place commitments get dropped