Did the Agent Really Do It? Reconstructing Asynchronous Tool Outcomes in MCP and Agent Platforms

The Callback Is Not the Outcome: Evidence-Based Verification of Asynchronous Tool Results for AI Agents Did the Agent Really Do It? Reconstructing Asynchronous Tool Outcomes in MCP and Agent Platforms Abstract: When an AI agent’s tool call says “completed,” did the intended action actually happen? Today, often nobody can say for sure. AI agents increasingly invoke tools asynchronously. A request returns a task handle, and the real outcome arrives later through polling, push notifications or webhooks, frequently after connection loss, server-instance failover, tool-catalog updates, cancellation, retries and concurrent quota consumption. In practice, a “completed” task status, a delivered callback, a subscription notification or a billing receipt is treated as proof that the intended action happened, on the intended resource, under the intended schema and within budget. None of these signals establishes that on its own. As agent platforms move toward long-running tasks, background execution and remote tool servers, this gap becomes a direct source of silent failures, duplicate actions and unexpected charges. This disclosure describes a vendor-neutral method for reconstructing, from independently sourced and provenance-qualified evidence, whether an asynchronous tool outcome actually corresponds to the original request. The method binds every task, callback and notification to a generation-qualified original invocation; determines the tool schema actually in force at execution time rather than the latest catalog entry; evaluates push events against the subscription filter the server acknowledged rather than the one requested; resolves resource identifiers by generation to detect reassignment; and reconstructs reservations, commits, retries, duplicate postings and credits to detect concurrent budget overruns. Results are combined using three-valued (strong Kleene) logic under an explicitly declared verification objective. A proven contradiction yields INCONSISTENT, complete supporting evidence yields CONSISTENT, and missing evidence yields UNRESOLVED instead of silent success. The result is a read-only reconstruction and is never by itself an authorization to execute, retry, cancel or charge. The document provides a 48-stage, six-phase engineering workflow with flowcharts, formal definitions, six detailed embodiments, 70 formal relations, 40 traceable engineering checkpoints, and a comparison with existing mechanisms including MCP Tasks, signed webhooks, idempotency keys, durable workflow engines, distributed tracing, resource versioning, schema registries, quota systems and saga compensation. A reproducible reference laboratory in Python, JavaScript and Go runs 71 shared test vectors, all passing. The laboratory is synthetic and self-authored; it is not peer reviewed and does not demonstrate production conformance. Industry relevance: the problem addressed applies directly to agent ecosystems built on the Model Context Protocol, introduced by Anthropic and since adopted across the industry, and to agent platforms offering background or long-running tasks, webhooks and tool integrations, such as those of Anthropic and OpenAI. This industry-relevance statement is AI-generated and does not assert that any named company uses, implements or infringes the disclosed method; readers must cross-check it against the original PCT filings before drawing any conclusion. This disclosure is based on the Mothership application PCT/IB2026/055615, filed 4 June 2026 and publicly disclosed 16 July 2026. The author is open to acquihire or strategic collaboration. Keywords: AI agent safety; asynchronous tool calls; Model Context Protocol (MCP); durable tasks; callback correlation; webhooks; idempotency; tool schema migration; subscription filters; push notifications; resource versioning; distributed usage accounting; quota reservation; three-valued logic; evidence provenance; TOCTOU; execution-effect verification; reproducible laboratory.