`Harness '${input.harness.harnessId}' could not execute host

Error message

`Harness '${input.harness.harnessId}' could not execute host tool '${toolCall.toolName}'.`

What it means

The SDK attempted to execute a host-side (custom, non-provider-executed) tool requested by the harness, but the execution wrapper reported `executed: false` — i.e. no tool executor ran for this tool call. This indicates the tool could not be dispatched (no matching executor available for `toolCall.toolName`) and the run cannot produce a tool result, so runPrompt throws.

Source

Thrown at packages/harness/src/agent/internal/run-prompt.ts:1083

                      result: preliminaryOutput as Extract<
                        HarnessV1StreamPart,
                        { type: 'tool-result' }
                      >['result'],
                    },
                    input.sessionWorkDir,
                  ) as Extract<HarnessV1StreamPart, { type: 'tool-result' }>;
                  result.enqueue({
                    type: 'tool-result',
                    toolCallId: toolCall.toolCallId,
                    toolName: toolCall.toolName,
                    input: undefined,
                    output: stripped.result,
                    preliminary: true,
                  } as TextStreamPart<TOOLS>);
                },
              });
              if (!execution.executed) {
                throw new Error(
                  `Harness '${input.harness.harnessId}' could not execute host tool '${toolCall.toolName}'.`,
                );
              }
              await telemetry.toolEnd(toolCall.toolCallId, execution.outcome);
            })(),
          );
        }
      }
      await waitForOutstandingHostToolExecutions();
      const isTurnSuspending = input.isTurnSuspending?.() === true;
      if (isTurnSuspending) {
        if (finalFinish == null) {
          /*
           * A timed slice may stop in the middle of a model step. Its partial
           * content remains in the bridge replay log for the next slice, but it
           * cannot form a valid StepResult in this slice because no finish-step
           * has arrived yet.
           */

View on GitHub (pinned to 69428b1f8b)

Solutions

Ensure every tool the harness can request is declared in the agent's `tools` with an executable `execute` implementation.
Verify tool names match exactly (case-sensitive) between harness configuration and the SDK toolset.
Check `builtinToolFiltering`/tool filters are not excluding a tool the harness still tries to call.
Keep the toolset stable across all steps of the run.

Example fix

// before
new HarnessAgent({ tools: { read_file: readFileTool } }); // harness calls 'bash'

// after
new HarnessAgent({ tools: { read_file: readFileTool, bash: bashTool } });

Defensive patterns

Strategy: validation

Validate before calling

for (const name of harnessRequestedToolNames) {
  if (!hasTool({ tools: activeTools, toolName: name })) {
    throw new Error(`Harness may request tool '${name}' which is not in the active toolset.`);
  }
}

Try / catch

try {
  await agent.run(...);
} catch (e) {
  if (e instanceof Error && e.message.includes('could not execute host tool')) {
    // add the missing tool or correct its name/filtering
  }
  throw e;
}

Prevention

Declare every tool the harness can invoke with an `execute` implementation.
Match tool names exactly (case-sensitive) between harness config and toolset.
Verify tool filters do not exclude harness-requested tools.

When it happens

Trigger: During host tool execution inside runPrompt, `execution.executed` is false after invoking the tool execution machinery — the tool name has no registered executor/implementation in the active toolset or execution was refused.

Common situations: The harness requests a tool that was filtered out or not declared in `tools`; typos or casing differences in tool names between harness config and SDK toolset; toolset swapped between steps so a previously available tool vanished; provider-executed vs host-executed flag mismatch.

Related errors

AI-assisted analysis of vercel/ai@69428b1f8b (2026-08-30). Data as JSON: /api/errors/34de920c18fe494c. Report an issue: GitHub.