You finished the quickstart. The my-investigator project on your laptop runs end-to-end. This page walks you through the local development loop you’ll use to extend it: hot-reload edits, trace inspection in Studio, and a durability test that kills the worker mid-run.

Time: about 10 minutes.

You’ll learn:

• Edit the workflow and see changes apply on the next run
• Read the trace for a paused run in Studio
• Kill the worker mid-pause, restart, and watch the workflow resume

Prerequisites:

• Completed the quickstart. You have my-investigator/ checked out and agnt5 dev in one terminal.

Step 1: Edit the workflow

Open src/agnt5_quickstart/workflows.py and tighten the prompt:

INVESTIGATOR_PROMPT = (
    "You investigate technical and operational questions for an engineering team. "
    "Use the DeepWiki MCP tools to read documentation and ask questions about "
    "GitHub repositories — that's your primary evidence source. If web search is "
    "available, use it sparingly to corroborate community signal. "
    "Separate first-party evidence (docs, source code) from public commentary. "
    "Return a concise brief: answer, evidence, risks, recommendation, open questions. "
    "Cite specific file paths and commit ranges from the source repo when relevant."  # NEW
)

Save the file. The terminal running agnt5 dev shows the worker reconnect:

File changed: src/agnt5_quickstart/workflows.py
Reloading components...
Registered components: investigate_with_review, save_report
Worker connected

No restart needed. The next run picks up the new prompt.

Step 2: Trigger another run and watch the trace

In your second terminal:

agnt5 run investigate_with_review --input '{
  "question": "Should we adopt Polars to replace Pandas in our analytics pipeline?"
}'

Open Studio (default https://app.agnt5.com; agnt5 context show if your context is custom). The new run shows up at the top of your project’s runs list. Click into it.

The trace shows:

• The workflow input (question).
• The MCP connect step against mcp.deepwiki.com/mcp.
• Each model call inside the agent loop, with input messages and output.
• Each tool call: DeepWiki read_wiki_structure, ask_question, plus any web_search_preview calls if you’re on the provider-hosted path. Built-in tool calls are marked built_in: true.
• The wait_for_user step, paused with the brief and the three options.

The trace is a record of every checkpointed boundary. There is no “agent black box” — every model call and every tool call is its own step.

Step 3: Verify durability

The HITL pause is the hard one. AGNT5 promises the workflow is not held in process memory. Verify it:

1. With the run still paused at review, switch to the terminal running agnt5 dev and stop it: Ctrl-C.
2. Wait 10 seconds. Confirm the worker is gone (ps aux | grep agnt5_quickstart returns nothing).
3. Restart it: agnt5 dev. Watch it reconnect and re-register investigate_with_review and save_report.
4. Open Studio and approve the brief.

The workflow resumes from the wait_for_user step. The agent does not re-call the model. The MCP server does not get re-queried. The save_report step runs and writes the file.

cat .agnt5/reports/*.md

The report contains the brief that was drafted before you killed the worker, with whatever edits Studio captured.

What that demonstrated

You exercised three properties that distinguish AGNT5 from a plain agent loop:

• Hot reload. Source edits register without a process bounce. The dev session is the development surface; you don’t redeploy locally.
• Glass-box trace. Every model call, tool call, and human-review pause is a discrete step in Studio. The trace is the artifact you’ll come back to when something is wrong.
• Durable pauses. A long-running pause (a human review, a webhook callback, a scheduled wait) is not a process. It’s a checkpoint. Workers come and go; the workflow does not.

These are the same properties that make the same workflow run unchanged in cloud. That’s the next page.

Next steps

• Run in cloud — promote the same workflow to a managed environment with agnt5 deploy.
• Workflows — the durable-execution model that makes the trace and the resume possible.
• Agents — the model→tool→model loop and how Agent composes with @workflow.