Skip to content

Latest commit

 

History

36 Commits

Folders and files

NameName
Last commit message
Last commit date
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

Debugging Agents in Different Environments

A one-hour workshop. Three AI agents run in three environments. Each can fail across several steps, so the final answer is only part of the story. You use your coding agent, with Sentry's MCP server and skills attached, to add agent tracing, verify it, find the failures, and set up the dashboards and alerts that matter for each critical path.

Nobody instruments by hand. Every step is a prompt. The workshop teaches what to ask for and how to check the result.

Three ways in

Your coding agent reaches Sentry in three ways. The labs use all three, and each lab names the one that did the work.

Way What it is Where the labs use it
MCP server mcp.sentry.dev. Tools such as search_issues, search_events, get_trace_details, analyze_issue_with_seer, and update_issue. The agent calls them while it works. Lab 3 finds the issue, pulls the trace and the replay, and triages it. Lab 6 triages the issue an alert opens.
Plugin skills sentry-instrument, sentry-debug-issue, sentry-create-alert, and others from https://github.com/getsentry/plugin-claude. Step-by-step instructions the agent follows on top of the MCP tools. Claude Code gets them from the sentry plugin. Other agents can copy them. Labs 2 and 4 instrument. Lab 3 debugs. Lab 6 creates alerts.
sentry CLI sentry issue list, view, explain, plan, resolve; sentry agent-conversation list, view; sentry alert; sentry dashboard create. Runs in any terminal and in CI, and the agent can call it too. Lab 3 shows the same issue from the terminal. Lab 6 creates the dashboard and lists the alerts.
App Environment Runtime You do
apps/storefront Browser chat on a Next.js app Next.js, Vercel AI SDK, Postgres on Neon Hands-on
apps/slack-agent Long-lived Slack bot Node, Bolt (Socket Mode), Vercel AI SDK, same Postgres Follow along
apps/pr-reviewer GitHub Actions job Node, Flue (OpenTelemetry), OpenRouter Follow along

All three call Claude Sonnet 5 through OpenRouter: the storefront and the Slack agent through the Vercel AI SDK, and the PR reviewer through Flue. The presenter shares one OpenRouter key at the start of the workshop, so no account is needed. Neon's instant Postgres needs no account either.

Before the workshop

Do these once, before the day. Each takes a few minutes. See labs/00-setup.md for the details.

  1. A Sentry account: https://sentry.io/signup/
  2. An OpenRouter API key: the presenter shares one at the start of the workshop, so no account is needed.
  3. A coding agent with Sentry's MCP server attached. Claude Code users install the sentry plugin. Other agents add https://mcp.sentry.dev/mcp.
  4. Node 22 and git.

Then:

git clone https://github.com/getsentry/debugging-agents-workshop
cd debugging-agents-workshop
./scripts/setup.sh

The script checks Node, creates a Postgres database on Neon, seeds it, and asks for your OpenRouter key and Sentry DSN.

Labs

# Lab Minutes
0 Setup before the day
1 Read a trace 5
2 Instrument the storefront by prompt 10
3 Debug the refund that fails 10
4 The same agent in Slack 10
5 The same agent in GitHub Actions 10
6 Alerts and dashboards for each critical path 10

Every prompt is also in PROMPTS.md, ready to paste.

Coding agents take different amounts of time and produce different diffs. Nobody waits for a prompt to finish. When the room moves on, apply the finished instrumentation for the current app and continue:

./scripts/solution.sh storefront    # or slack-agent, pr-reviewer

solutions/<lab>/ holds each patch and a README that lists what it adds, so you can compare it with what your agent did.

Reference

The fully instrumented versions of these apps live at https://github.com/getsentry/sentry-agent-tracing-examples.

The same Slack assistant also exists on Vercel's eve agent framework, in slack-agent-eve and slack-agent-eve-otel in that repo. eve receives Slack events over a webhook and needs a Vercel deploy, so it is not used in the labs. It shows what changes when a framework owns the spans: one version uses the Sentry SDK inside eve's OpenTelemetry pipeline, the other exports OTLP straight to Sentry.

License

MIT. See LICENSE. The storefront is built on Vercel's Next.js Commerce template, and its MIT license is in apps/storefront/license.md.

About

Workshop: Debugging Agents in different environments. Trace an ecommerce chatbot, a Slack agent, and a PR review agent with Sentry.

Topics

Resources

Code of conduct

Security policy

Stars

0 stars

Watchers

0 watching

Forks

Releases

Packages

Used by

Contributors

Languages