Agentium exposes 5 tools with basic schemas and descriptions. All tools have descriptions (10-200 chars) and input schemas with typed parameters. However, descriptions are minimal and lack LLM-optimization guidance (e.g., 'when to use', 'what it returns'). Parameter descriptions are present but terse. No output schemas documented. Error handling is absent, no recovery guidance, no categorization of retryable vs fatal errors. Tools lack composition clarity: transfer_to_agent and complete are control-flow tools (not domain tools), while artifact tools form a coherent subsystem. No tool annotations (readOnlyHint/destructiveHint) despite clear risk profiles. Naming is clear (verb_noun pattern) but descriptions do not explain prerequisites or dependencies.

• Expand tool descriptions to 50-200 chars. For transfer_to_agent: 'Transfer the conversation to a specialist agent when you cannot handle the request. Provide the agent name and reason. Available agents: [list]. Returns confirmation or error if agent unavailable.'
• Document output schemas for all tools. E.g., storeArtifact returns {pointer: string, size: number, expiresAt?: string}. listArtifacts returns {artifacts: [{pointer, name, size, preview, contentType, createdAt}], total: number}.
• Add error handling guidance. E.g., getArtifact: 'If pointer not found, returns error with available artifact names. Call listArtifacts to discover valid pointers.'
• Add tool annotations to schema. Mark storeArtifact with destructiveHint (modifies state), getArtifact and listArtifacts with readOnlyHint. Include idempotentHint for getArtifact.
• Clarify parameter constraints. storeArtifact.value: 'String or JSON.stringify(object). Max 10MB. If larger, split across multiple artifacts.' getArtifact.pointerOrName: 'Prefer pointer (art:...) for speed; name lookup is slower.'

Inferred effective spec: 2026-07-28+.

• Current-spec patterns present: Streamable HTTP