Agentic Builder Path / Agentic Systems
Skills, Tools and Harnesses
Understand the building blocks that let agents act safely and repeatably.
Understand the building blocks that let agents act safely and repeatably.
Use the brief to sharpen a real ai upskill conversation: what is the decision, what evidence matters, and what should remain human-led?
Capture one design rule you would reuse when reviewing an AI workload, assistant, or operating model.
Executive note
Core Idea
Once an AI system can use tools, it stops being only a text generator. It can retrieve information, write records, call APIs, run checks, create tickets, query databases, or trigger workflows.
That is where agentic AI becomes useful. It is also where it becomes harder to govern. A model answer can be wrong; a tool action can change a system.
Section 2 of 7
Skill
A skill is a packaged way of doing a task. It can include instructions, examples, constraints, tool choices, and expected outputs.
For example, a "summarize incident" skill might tell the agent to read logs, separate facts from assumptions, list customer impact, include timestamps, and finish with open questions.
Treat skills like delivery assets:
- Version them.
- Review them.
- Test them with examples.
- Retire skills that drift away from current practice.
Section 3 of 7
Tool
A tool is a capability the AI application can call. It may search content, read a ticket, update a record, fetch a price, generate a chart, or call another service.
The tool contract should be boring in the best possible way: clear input, clear output, explicit permissions, predictable errors, and telemetry.
Poor tool design makes the agent guess. Good tool design reduces guessing.
Section 4 of 7
Harness
A harness is the runtime wrapper around the model, skills, tools, identity, policy, logging, and evaluation.
The harness decides what the agent can see, what it can call, how it handles failures, when it must ask for approval, and what evidence is recorded.
In enterprise use, the harness is often more important than the model. It is where safety and repeatability live.
Section 5 of 7
A Practical Pattern
Use this structure when designing agent capabilities:
1. Define the task the user is trying to complete.
2. Identify the minimum information needed.
3. Select only the tools required for that task.
4. Add skill instructions that encode good practice.
5. Add approval gates for high-impact actions.
6. Capture telemetry and audit evidence.
Section 6 of 7
Recommended Practices
- Prefer small, task-specific tools over one large "do anything" tool.
- Give tools names that describe business intent, not implementation detail.
- Keep sensitive operations behind explicit authorization checks.
- Make tool outputs structured so the model can use them reliably.
- Evaluate the full workflow, not only the model answer.
Section 7 of 7
Remember This
Skills teach the system how to behave. Tools let it act. Harnesses make the behavior controlled, observable, and repeatable.
If you cannot explain which tools an agent can call and why, the workload is not ready for serious use.