Role-Specific Harness Engineering and Token Optimization
Same model, different harness, very different results. Participants build a harness tuned to their role and measure what it saves on a real task.
- Audience
- Engineers and tech leads already using skills, rules and MCP. Run per role (dev, QA, system engineering).
- Duration
- 3 to 3.5 hours, hands-on throughout
- Group size
- 15 to 20 participants, on their own laptops with the tools they already use
- Hands-on
- On the team's real tasks, or on prepared exercises I bring
- Includes
- A tailoring call with the group lead beforehand and a follow-up session about 30 days later
- Where
- On-site in Israel, on-site abroad (I fly in for a few days and can run several sessions or teams back to back), or remote
What we cover
- Agent = model + harness, and why a developer's harness is the wrong one for QA.
- Anatomy of a harness: guides (AGENTS.md, rules, skills) vs. sensors (tests, hooks, reviewer agents).
- Where tokens actually go, and the levers that cut them: progressive disclosure, MCP vs. CLI, model and reasoning level per step, sub-agents for noisy work.
- Keeping the harness alive: versioning, ownership, sharing it across the team.
What participants do
- 1 Audit their own setup and find what loads before the first prompt.
- 2 Build a harness for their role on a real team task.
- 3 Run the task with and without it and compare cost and quality.
What they take home
- A working, versioned role harness in their repo.
- A before/after measurement on a real task.
- A repo with the workshop's exercises and materials.
Frequently Asked Questions
Can you tailor it to our team?
Yes. Before the session I have a tailoring call with the group lead. The hands-on part then runs on the team's real tasks, or on prepared exercises I bring.
Does it matter what language or stack we use?
No. The agent reads and writes the code, whatever the language. Participants work on their own laptops with the tools they already use. Exercises run on the team's real tasks, or on prepared exercises I bring.
Can you run it on-site abroad?
Yes. I fly in for a few days and can run several sessions or teams back to back. It also runs on-site in Israel or remote.
What happens after the workshop?
About 30 days later I run a follow-up session with the group. Participants also keep a repo with the workshop's exercises and materials.
Related workshops
Advanced Skills: From "It Worked Once" to Reliable
For the skill you built, got disappointed by, and never used again. Why skills fail, how to fix them, and how to prove one helps before sharing it.
See the Agent Skills workshopAgent Orchestration and Automation with Codex
Several agents in parallel can save real time, or just burn tokens. Participants learn to tell the difference, then take a multi-agent workflow off the laptop and run it automatically.
See the Agent Orchestration workshopGet in touch
Tell me about your team and what you want them to walk away with.