isError 플래그

When an MCP tool runs, two things can happen: it succeeds, or it fails. The model needs to know which — clearly and unambiguously — to decide what to do next.

If a failure looks like a normal result, the agent may treat garbage as truth, hallucinate a recovery, or silently move on. The fix is a dedicated failure signal: the isError flag.

In this lesson you'll learn how to signal failure cleanly so the agentic loop can route intelligently instead of guessing.

A tool result carries an isError boolean. When isError is true, you're telling Claude: this tool did not produce a valid result — treat the content as a failure report, not data.

This is structurally separate from your tool's normal output. The model can branch on it without parsing prose: success path vs. failure path. That separation is the whole point.

The classic anti-pattern is a generic error string like "Operation failed". It tells the model that something went wrong but nothing it can act on.

Can it retry? Was the input malformed? Did the user lack permission? Is there partial data to salvage? A generic message answers none of these — so the agent stalls or improvises badly.

Generic errors block recovery. Structured errors enable intelligent routing.

A well-formed MCP error pairs isError: true with a structured body. The exam-standard fields are:

• errorCategory — one of transient, validation, business, permission
• isRetryable — can the same call succeed if tried again?
• message — human-readable explanation
• attempted_query — exactly what the tool tried to do
• partial_results — anything usable it managed to gather

Together these let the model decide: retry, reformulate, escalate, or proceed with partial data.

The four categories aren't decoration — each implies a different next action:

• transient — temporary fault (timeout, rate limit). Usually retryable; recover locally.
• validation — bad input. Don't blind-retry; fix the arguments first.
• business — a rule was violated (e.g. refund exceeds policy). Often needs escalation, not retry.
• permission — caller lacks access. Retrying won't help; escalate or request credentials.

The category turns a vague failure into a routing instruction the model can follow.

Whether a failure is worth retrying is often invisible from the message text alone. Make it explicit with isRetryable.

A timeout (transient) is retryable. A malformed argument (validation) is not — retrying the same bad input just fails again. A permission denial is not retryable without new credentials.

By stating isRetryable directly, you keep retry decisions deterministic instead of leaving them to probabilistic text-reading.

When the model decides how to recover, it needs to know what was actually tried. Including attempted_query means the agent can reformulate intelligently instead of repeating the same failing call.

This is part of good error propagation: structured context = failure type, attempted query, partial results, and alternatives. The richer the context, the better the recovery routing.

A tool can fail and still have gathered something useful. A multi-source lookup might return 3 of 5 records before the 4th source times out.

Returning partial_results alongside the error lets the agent proceed with what it has, annotate the gap, and avoid restarting from zero. Discarding partial data on any failure wastes work and degrades answers.

A critical distinction: an access failure is not the same as a valid empty result.

• isError: true — the tool couldn't complete (timeout, denied, bad input). Maybe retry or escalate.
• isError: false with empty content — the tool ran fine and the honest answer is "no matches."

Conflating these is a common bug: an empty search marked as an error triggers pointless retries, while a real failure marked as empty hides the problem. Keep them distinct.

The structured signal drives where recovery happens. In a multi-agent system, a subagent should recover transient faults locally — retry the timeout, re-issue the call — and only bubble up failures it truly can't resolve.

When it does escalate, it passes the structured error with partial results so the coordinator can decide: route elsewhere, ask for more identifiers, or surface the gap. Never silently suppress a failure, and never abort the whole workflow over one recoverable fault.

Inside an MCP tool handler, wrap the work and return a structured error on failure instead of letting an exception leak as a generic string.

Notice how each branch sets isError, a category, and a retry hint — giving the agentic loop everything it needs to route the next step deterministically.

A subagent's MCP tool queries a customer's order history. One of three backend shards is unreachable; the other two return 8 orders. What should the tool return?

Key takeaways for the isError flag:

• isError: true is the structural signal that a tool did not produce valid data — separate from normal output.
• Pair it with errorCategory (transient / validation / business / permission), isRetryable, message, attempted_query, and partial_results.
• Generic errors like "Operation failed" block recovery; structured errors enable intelligent routing.
• Distinguish an access failure from a valid empty result — never mark "no matches" as an error.
• Recover transient faults locally; escalate non-recoverable ones with partial results. Avoid silent suppression and avoid aborting the whole workflow on one fault.