บริบทข้อผิดพลาดแบบมีโครงสร้าง

In a multi-agent system, a subagent will eventually hit a failure: a database is down, a query returns nothing, a permission is denied. How that failure is reported decides whether the coordinator can recover intelligently or just gives up.

A generic status like "Operation failed" blocks recovery — the coordinator has no idea what to do next. A structured error context turns a dead end into a routing decision.

This lesson covers the four pillars of a good error context: failure type, attempted query, partial results, and alternatives.

Compare two error payloads coming back from a subagent or tool.

The generic version tells the coordinator nothing actionable. It can't decide whether to retry, ask the user for more input, or escalate. Silent suppression is even worse — the workflow continues as if data exists when it doesn't.

The structured version names what failed and why, which is the first step toward an intelligent next move.

The first job of an error context is to classify the failure. Structured MCP errors carry an errorCategory field with a small, fixed vocabulary:

• transient — temporary infrastructure fault (timeout, rate limit). Often retryable.
• validation — the input was malformed.
• business — a domain rule blocked the action.
• permission — access was denied.

The companion isRetryable boolean removes guesswork: the coordinator reads it directly instead of inferring intent from a free-text message.

One distinction trips up architects constantly: an access failure is not the same as a valid empty result.

• Failure: the lookup couldn't run — DB unreachable, permission denied. This might be worth a retry.
• Empty: the lookup ran successfully and found zero matches. Retrying changes nothing — the answer is genuinely "none".

Conflating the two leads to pointless retry loops on empty results, or to treating a real outage as "no data found". Always model them as separate states.

The coordinator did not run the failing operation itself, so it can't see what was tried. Include the attempted_query verbatim in the error context.

This serves two purposes:

• It lets the coordinator decide whether to retry with the same query or reformulate it (e.g. broaden a filter that was too narrow).
• It gives the human reviewer, on escalation, the exact reproduction case instead of a vague "search failed".

A failure rarely means zero work got done. A research subagent may have gathered 6 of 10 sources before a provider rate-limited it. Throwing all of that away — or aborting the whole workflow — wastes real progress.

Attach whatever was successfully collected as partial_results. The coordinator can then aggregate what exists, annotate the gap, and decide if the remainder is worth another attempt.

Never silently suppress the failure and present partials as if they were complete.

The most useful error contexts don't just describe the wall — they point at a door. The alternatives field suggests concrete next moves the coordinator (or a human) can take.

Examples: "retry against the read replica", "broaden the date filter", "ask the user for an order number", "escalate to a human with the partial results attached".

This is what turns a structured error from a report into a routing instruction.

Structured context drives a clear policy. Handle transient faults inside the subagent — retry the timeout, back off the rate limit — so the coordinator never even sees a recoverable blip.

Only when a fault is genuinely non-recoverable (policy cap hit, permission denied, retries exhausted) do you escalate upward — and you escalate with the partial results and alternatives attached, not as a bare "failed".

The goal: don't abort the entire workflow because one branch failed.

Free-form error dicts drift over time. Enforce the contract with a JSON Schema via a tool / structured output so the subagent must populate the right fields.

Key rule from structured-output design: mark a field required only if it is always present. errorCategory and message are always there — require them. partial_results and alternatives may be absent — leave them optional, or the model will fabricate them to satisfy the schema.

Subagents do not inherit the coordinator's conversation history. So when a subagent fails, the coordinator only knows what the error payload explicitly carries.

That is exactly why attempted_query and partial_results must be in the structured error — there's no shared memory the coordinator can fall back on. The error context is the entire bridge between the two.

Trim it to the relevant fields, but never strip out the four pillars.

A production-grade error context lets the coordinator act without re-running anything itself:

• failure type + isRetryable → retry vs. escalate decision
• attempted query → reproduce or reformulate
• partial results → salvage progress, annotate the gap
• alternatives → the concrete next move

This is the difference between a brittle pipeline that dies on the first hiccup and a resilient system that degrades gracefully and routes around damage.

A research subagent was asked to gather 10 sources. After collecting 6, the news API rate-limited it (HTTP 429). What should the subagent return to the coordinator?

Key takeaways:

• Generic errors block recovery; structured errors enable intelligent routing.
• Always carry the four pillars: failure type, attempted query, partial results, alternatives.
• Use errorCategory (transient / validation / business / permission) + isRetryable to drive the retry-vs-escalate decision.
• Distinguish an access failure (maybe retry) from a valid empty result (no matches — retrying won't help).
• Recover transient faults locally in the subagent; escalate non-recoverable ones with partials attached.
• Never silently suppress errors and never abort the whole workflow on a single failure.
• Enforce the shape with a schema, but only require fields that are always present — optional pillars must stay optional to avoid fabrication.