TechPrufer
Get started
Temporal
Temporal logoTemporal · Feature trail

AI-Assisted Workflow Debugging

2 posts · May 2026 – Aug 2026

What this trail covers

Temporal MCP server provides AI assistants with read-only access to Temporal Cloud and self-hosted clusters, enabling natural language querying of workflow status, history, and infrastructure. This facilitates debugging by correlating runtime state with code, improving cross-team troubleshooting. This post details the integration of the Temporal MCP server with AI editors, allowing engineers to query workflow status, history, and infrastructure using natural language directly within their development environment. Coinbase built their own read-only Temporal MCP server to integrate with their AI editor, enabling engineers to ask natural language questions about workflow failures and get answers directly in their IDE. The system uses FastMCP and exposes tools for workflow operations, schedule operations, and infrastructure inspection. It communicates over stdio using the MCP protocol and handles namespace resolution, auth, and payload decoding. The example demonstrates how a single question can trigger multiple tool calls to diagnose a workflow failure.

All Temporal trails
  1. Label your agent stepsThis post introduces the ability to set human-readable summaries on Temporal Activities, Timers, and Workflows. For Activities, this is achieved via the `summary` option during execution. For Timers, the `summary` field can be set during `workflow.sleep`. For Workflows, `static_summary`, `static_details`, and `current_details` fields are introduced. `static_summary` and `static_details` are immutable and set at workflow start, while `current_details` is mutable and can be updated during workflow execution. These features enhance the observability of agentic workflows by providing clear labels in the Temporal UI's timeline and compact history views, making it easier to understand the state and progress of complex agentic loops and debugging sessions.
  2. How Sherlocks AI uses Temporal to orchestrate AI agents for incident resolutionThis post details how Sherlocks AI uses Temporal to orchestrate AI agents for incident resolution. It highlights the use of Temporal Workflows for four distinct patterns: Knowledge Graph construction, Infrastructure scanning, AI agent investigations, and Event ingestion. The authors emphasize Temporal's Durable Execution, Retry Policies, Scheduling, Human-in-the-loop with Signals, Observability, and model agnosticism as key decision factors. They also describe structural choices like per-tenant Workflows, Workers as separate processes, and using a Codec Server for encryption. The post details how Temporal replaced previous solutions like threads, message queues, and agent frameworks, enabling faster agent improvement and reducing operational overhead.
  3. How Kelet AI built a durable agent that debugs durable agents.This post details Kelet AI's architecture for building a durable agent that debugs other AI agents, specifically focusing on identifying cross-session quality failures. It highlights the necessity of durable orchestration over simple agent loops for AI quality analysis. The architecture uses a four-level Temporal Workflow hierarchy: Session Workflow (with a 5-minute silence window debounce), Signal Workflows for parallel enrichment, Agent Aggregation Workflow to collect hypotheses across sessions, and Investigate Issue Workflow for root cause analysis and prompt patching. It emphasizes Temporal primitives like long-running state, durable execution, event-driven coordination, and cross-workflow gating. The integration with Temporal involves a plugin that configures OpenTelemetry and propagates session attributes via Temporal headers for deterministic replay and `continue_as_new` survival. The post also details the filtering of self-monitoring workflows to prevent infinite loops.
  4. Bringing Temporal into your AI editor: How Coinbase debugs Workflows with MCPCoinbase details the construction and usage of their internal read-only Temporal MCP server, integrated with their AI editor. This system allows engineers to query Temporal Cloud and self-hosted clusters using natural language, facilitating debugging by correlating runtime state with code. The post highlights the use of FastMCP, the `@mcp.tool()` decorator for defining tools like `list_workflows` and `get_workflow_history`, and the communication protocol (JSON-RPC over stdin/stdout). It showcases a real-world debugging scenario where an engineer asks a question, the AI assistant uses multiple tools to gather information (workflow list, then history), and provides a root-cause diagnosis directly within the editor. The contribution emphasizes the power of combining runtime state with source code context for cross-team troubleshooting.