AgentRunResult

AgentRunResult<Tools, TOutput> = object

Defined in: packages/core/src/agent/types.ts:272

Result of one agent.run / agent.stream turn (possibly several model requests).

text / output: the final model response. Intermediate tool-call requests are in messages / newMessages and as agent_tool_call / agent_tool_result events — not in text. A tool whose execute returns an AsyncIterable (streaming progress, e.g. a long-running shell command) emits an agent_tool_result per yielded value plus one more (see that event’s own doc comment), each preliminary: true except the last — none of the preliminary ones reach messages / newMessages or the model itself, only the final one does.

output: typed payload from AgentDefinition.outputSchema (or a per-call override). When no schema is set, this is the same string as text. When a schema is set, this is the AI SDK’s complete parse of the final answer text. A schema plus tools does not send a JSON response format on the tool loop (that combination makes some models skip the tools and emit a minimal object). If that loop stops with text that is not schema JSON, one tool-free follow-up call parses into output, and text is that follow-up. The run rejects with NoObjectGeneratedError — raw model text on the error’s text field, and on agent_failed — when the answer is not JSON matching the schema. Zod field defaults apply only to a JSON object that parsed; they do not turn prose or a partial parse into a successful result. A stopWhen that ends on a tool call fails the same way and does not run the follow-up.