vercel/ai · error · Error

Deep Agents runtime was not initialized

Error message

Deep Agents runtime was not initialized

What it means

`runTurn` reads the module-level `agent` variable after a possible rebuild and throws if it is still undefined. Since `rebuildAgent` is true whenever `agent == null`, reaching this throw means agent creation silently failed to assign the variable or the code path was bypassed. It is a defensive invariant check that the Deep Agents runtime exists before streaming.

Source

Thrown at packages/harness-deepagents/src/bridge/index.ts:296

    agent = createDeepAgent({
      tools: [...mcpTools, ...hostTools],
      backend: createLocalShellBackend({ rootDir: workdir }),
      systemPrompt: start.instructions
        ? { suffix: start.instructions }
        : undefined,
      // Native skills loaded from the source dirs ($HOME-materialized + <workDir> for repo-provided skills).
      ...(start.skillsPaths?.length ? { skills: start.skillsPaths } : {}),
      ...(middleware.length > 0 ? { middleware } : {}),
      // Gate built-in tools behind HITL approval when the permission mode requires it.
      ...(interruptOn ? { interruptOn } : {}),
      // Real instance (LangGraph rejects `true` for root graphs); gives multi-turn memory.
      checkpointer,
    });
    agentConfigurationSignature = nextAgentConfigurationSignature;
  }
  const activeAgent = agent;
  if (activeAgent == null) {
    throw new Error('Deep Agents runtime was not initialized');
  }

  const hostToolNames = new Set((start.tools ?? []).map(t => t.name));
  const streamEventState = createDeepAgentsStreamEventState();
  const emitStreamEvent = createEmitStreamEvent({
    state: streamEventState,
    configuredModel: activeModel,
    hostToolNames,
    mcpToolNames,
    structuredOutputToolNames: new Set(
      currentResponseFormat?.map(format => format.name) ?? [],
    ),
    emit,
  });

  // After a stream segment ends, return the tool calls paused by HITL interrupts (empty when the turn is truly done).
  const readPendingApprovals = async () => {
    try {

View on GitHub (pinned to 69428b1f8b)

Solutions

1. Restart the bridge process (start a fresh harness session) so the agent is rebuilt from the start message.
2. Check logs above this error for a swallowed exception during `createDeepAgent` (e.g. MCP tool loading or middleware creation failures).
3. Avoid running turns concurrently against the same bridge process; turns share the module-level `agent`.
4. If reproducible, file a bug — this throw is an invariant guard and should be unreachable in normal operation.

try {
  await runTurn(start);
} catch (error) {
  if (error instanceof Error && error.message.includes('runtime was not initialized')) {
    // restart the bridge process and retry once
  }
  throw error;
}

When it happens

Trigger: `runTurn` reaching the streaming phase with `agent` still undefined — practically only if agent construction (`createDeepAgent`) was skipped or an exception path left the variable unset while execution continued.

Common situations: A corrupted bridge process state after a failed rebuild; race between concurrent turns mutating the module-level `agent`; a bug where the rebuild branch was not entered despite a null agent.

Related errors

• no active turn
• Invalid argument for parameter requests: requests must not b
• Invalid argument for parameter requests: request IDs must no
• Invalid argument for parameter requests: request IDs must be
• Invalid argument for parameter batch: batch must be a suppor

AI-assisted analysis of vercel/ai@69428b1f8b (2026-08-30). Data as JSON: /api/errors/3b9fb5468a6902b1. Report an issue: GitHub.