ODSC 2026: Engineering the Harness
This is the participant path for the two-hour ODSC workshop. You will not spend the session filling in starter code. You will make design decisions, compare them with a neighbor, and inspect what those decisions change in a real coding-agent trace.
Follow Get started for the workshop to install Canvas, connect a model, and add OpenCode or Pi through ACP. Open Exercise prompts for direct links to the tasks and templates. You can also follow the instructor demonstrations and complete the worksheet.
What You Will Build
You will leave with an annotated harness design covering:
- Trace and loop observability
- Retrieval and codebase understanding
- Context, memory, and skills
- Long-running execution state
- Tools, verification, and safety boundaries
- Model and reasoning routing
- Multi-agent architecture
How Each Exercise Works
- Inspect a problem or changed constraint.
- Choose among plausible harness techniques.
- Discuss your choice with a neighbor.
- Predict what the trace or measured result will show.
- Compare your prediction with the evidence.
- Mark the claim as supported, supported with a narrower condition, or not supported in this test.
- Record the decision and its tradeoff in the workshop worksheet.
Agent Canvas is our microscope. The instructor will use it to show conversation history, model decisions, tool calls, observations, edits, verification, compaction, and final answers in order.
Workshop Route
| Exercise | Question |
|---|---|
| Same Model, Different Harnesses | How do calls, context, time, and evidence change when the model stays fixed? |
| Do You Need an MCP? | Does MCP improve the result over an API the terminal can already reach? |
| What Should the Agent Remember? | Which instructions, skills, and memories earn their cost on future tasks? |
| Use the Strongest Model for Everything? | Can verification and escalation make routing cheaper without lowering quality? |
| Let It Run Until It Figures It Out? | Does more time create progress, and what must the harness still enforce? |
| What You Want Is a Multi-Agent System? | Does another agent add useful parallel work or coordination cost? |
After the Workshop
The optional long-project lab contains the exact incident benchmark prompt, reproduction steps, and guidance for adapting it to your own application.
The full self-paced course contains the runnable starter and solution projects behind the other exercises. Begin with P01: Agent Trace, then select projects based on the parts of your harness you want to investigate.