Skip to main content

Exploratory: Domain Agent Stretch Projects

These are exploratory stretch projects — not required to complete Module 10. They extend the domain agent build into more advanced scenarios.

Project 1: Second-Track Agent​

Estimated time: 60 minutes Extends: Module 10 lab (your primary track) Prerequisites: Primary track agent complete and passing evaluation checklist

What You Will Build​

Build a second domain agent using a different track from the one you completed in the lab. If you built Track A (DB Health), build Track C (Kubernetes Health) or Track B (FinOps). Use the reference SOUL.md and config.yaml templates from the Reference page as starting points.

Challenge​

Each track has a different evaluation paradigm:

  • Track A diagnoses point-in-time state against known thresholds
  • Track B analyzes time-series trends against a baseline
  • Track C traverses hierarchical state to identify cascading failures

The challenge is noticing how the skill design changes depending on what "good" means for each domain. A DB health agent needs precise threshold definitions. A FinOps agent needs baseline definition and anomaly significance scoring. A K8s agent needs depth-first traversal with noise filtering.

Steps​

  1. Read the lab guide for your second track in the Hermes repository
  2. Build the SOUL.md and config.yaml using the Reference page templates as starting points
  3. Write or adapt the primary skill for the second track (or use the provided reference skill)
  4. Run all simulated scenarios and apply the track-specific evaluation checklist
  5. Compare to your first track: what did you design differently and why?

Expected Deliverable​

A working second-track agent profile in a separate directory (my-track-b-agent/ or similar), passing all simulated scenarios. A brief comparison note: what you did differently in skill design for the second track.

Project 2: Cross-Track Incident Simulation​

Estimated time: 45 minutes Extends: Module 10 lab (requires two agents from different tracks) Prerequisites: Two agent profiles built (your primary track + Project 1, or two simulated agents)

What You Will Build​

A cross-domain incident scenario where both your agents contribute to a diagnosis. Example: "EC2 costs spiked AND database latency increased — determine whether they are related."

This simulates the real-world situation where a single incident has multiple observable manifestations across domains, and no single agent sees the full picture.

Challenge​

The challenge is the correlation step. Each agent will produce a correct domain-specific diagnosis. The integration challenge: how do you get from "FinOps agent found cost anomaly on us-east-1 EC2" + "DB agent found latency spike on RDS instance in us-east-1" to "hypothesis: these are caused by the same event — possible EC2 type change or storage migration"?

Steps​

  1. Craft a simulated incident using mock data that has correlated signals across two domains (example: CPU spike in CloudWatch that appears in both EC2 utilization AND RDS performance metrics at the same timestamp)

  2. Run each agent independently against the simulated data. Record each agent's diagnosis independently.

  3. Manually synthesize: given both diagnoses, what is the cross-domain hypothesis? What additional data would confirm it?

  4. Write the correlation logic as a cross-domain SKILL.md (from Module 7 Project 1) and test whether a single agent with both data sources can reach the cross-domain conclusion.

Expected Deliverable​

Two independent diagnoses + a written cross-domain analysis + a draft cross-domain skill that addresses the correlation step.

Project 3: Agent Comparison Study​

Estimated time: 30 minutes Extends: Module 10 lab (any track) Prerequisites: Your domain agent complete

What You Will Build​

A structured comparison of your agent's output vs. a human engineer's diagnosis for the same simulated scenario. This is how you measure whether the agent is actually useful — not by whether it runs, but by whether its diagnosis matches (or exceeds) what an expert would produce in the same time.

Challenge​

Bias in self-evaluation. It is easy to unconsciously design the human comparison to favor your own agent. The challenge is honest evaluation: where does the agent produce better output (faster, more systematic, catches things humans miss)? Where does it produce worse output (wrong confidence, misses context, overly verbose)?

Steps​

  1. Choose your most complex simulated scenario

  2. Human first: Without running the agent, write what you would diagnose given the same simulated data. Time yourself (aim for 10 minutes). Write the same structured output: Summary, Evidence, Root Cause, Recommended Actions, Escalation Decision.

  3. Agent second: Run the agent against the same scenario. Record the output verbatim.

  4. Comparison: For each output section, score 1-3: human better (1), equivalent (2), agent better (3). Write one sentence explaining each score.

  5. Improvement list: For each section where the human output was better (score 1), identify whether the gap is fixable with a skill update, or whether it requires human judgment that cannot be encoded.

Expected Deliverable​

Side-by-side comparison table + improvement list with at least one concrete skill update that would close a specific gap.

Which Project Should You Do?​

Your InterestRecommended Project
Building coverage across domainsProject 1 (second track)
Multi-agent coordination (preview of Module 11)Project 2 (cross-track incident)
Rigorous quality evaluationProject 3 (comparison study)
Under 30 minutes availableProject 3 — highest signal-to-effort ratio

All three projects prepare you for Module 11 (fleet orchestration), where multiple agents coordinate on cross-domain incidents. Project 2 is a preview of exactly that capability.

Last updated on