Incident management · Production

Fuzzy Labs builds an autonomous SRE agent using FastMCP and Claude

The problem

SRE teams spend significant time manually chasing root causes of production incidents by sifting through logs, inspecting Kubernetes services, and hunting for errors before communicating findings to the wider team.

First attempt

Relying on Claude Desktop and Cursor as MCP clients was insufficient for fully autonomous workflows: Anthropic gate-kept token usage and the tools required users to accept tool calls on the agent's behalf, preventing full autonomy.

Workflow diagram · grounded in source
1
CloudWatch detects error
trigger
“AWS CloudWatch detects the CRITICAL 500 error and alerts the agent”
2
LLM plans tool calls
ai_action
“The LLM plans its actions, starting with list_pods to locate the cart-service pod”
3
Retrieve Kubernetes logs
integration
“Next, get_logs retrieves the pod logs”
4
Analyze logs and identify culprit
ai_action
“The LLM analyses the logs and identifies RedisCartStore.cs as the culprit”
5
Fetch source file from GitHub
integration
“It fetches the file from GitHub for additional context and drafts a fix”
6
Post diagnosis to Slack
output
“Finally, it posts a diagnosis and recommendation in Slack”
Reported outcome

The team implemented a custom MCP client running a fully autonomous SRE agent that diagnoses production incidents end-to-end and posts findings to Slack; tool caching reduced cost per diagnosis by 83%.

Reported metrics
Cost per diagnosis83%
Reported stack
FastMCPClaudeGitHub MCP ServerSlack MCP ServerCloudWatchSlackGitHubKubernetes
Source
https://www.fuzzylabs.ai/blog-post/how-we-built-our-sre-agent-using-fastmcp
Read source ↗

Frequently asked questions

What did this team achieve with this AI workflow?

The team implemented a custom MCP client running a fully autonomous SRE agent that diagnoses production incidents end-to-end and posts findings to Slack; tool caching reduced cost per diagnosis by 83%.

What tools did this team use?

FastMCP, Claude, GitHub MCP Server, Slack MCP Server, CloudWatch, Slack, GitHub, Kubernetes.

What results were reported?

Cost per diagnosis: 83% (source-reported, not independently verified).

What failed first in this deployment?

Relying on Claude Desktop and Cursor as MCP clients was insufficient for fully autonomous workflows: Anthropic gate-kept token usage and the tools required users to accept tool calls on the agent's behalf, preventing…

How is this incident management AI workflow structured?

CloudWatch detects error → LLM plans tool calls → Retrieve Kubernetes logs → Analyze logs and identify culprit → Fetch source file from GitHub → Post diagnosis to Slack.

Related incident management cases