A shared definition
Agentic engineering,
plainly.
It is the work of designing software where AI can take bounded, multi-step action toward a goal. The model matters, but so do the conditions around it.
Five principles
A real need
Start with a person, a decision, and a useful outcome.
Relevant context
Give the work the few rules and sources that matter.
Bounded action
Let the agent use clear tools in a safe space.
Visible evidence
Show the plan, result, checks, and what changed.
Human judgment
Leave room to clarify, redirect, review, or approve.
Glossary
- Agent
- A model working through a task over several steps with tools and feedback, not just one reply.
- Context engineering
- Selecting the right project history, constraints, and source material for the work at hand.
- Harness
- The workbench around an agent: tools, boundaries, memory, checks, and decision points.
- Evaluation
- Evidence that the system did the useful thing, not merely something convincing.
A closer look
The harness atlas
Four diagrams that show how I think about AI work: give it context, keep it understandable, check the result, and leave a useful trail for the next person.
Scroll sideways to inspect each drawing.
Plate 01 · Around the model
The model is one part of the system.
It also needs memory, tools, clear boundaries, ways to check its work, and a person who can stay in control.
- A model should not get to choose its own boundaries.
- A person should be able to see what the system did.
Plate 02 · Shared work
Uncertainty needs conversation.
When a request is unclear, it helps to investigate, ask questions, make a plan, and look at it together before the costly work begins.
- An unclear detail is a reason to ask, not to guess.
- It is much easier to change a plan than rebuild the wrong thing.
Plate 03 · Shared knowledge
Make a useful idea easy to pass on.
People have always used writing, records, and tools to share what they know. AI systems also need memory and reusable instructions, but they should not hide the human choices behind them.
- Write down enough for another person to understand the idea.
- Reuse what works without erasing who made the decision.
Plate 04 · Staying involved
Keep people present at the important moments.
A plain-language request can become a plan, a manageable change, something visible to review, and then a real choice about what happens next.
- People can clarify, redirect, review, and approve along the way.
- A checked result is more useful than a confident answer.
These diagrams are how I explain the ideas today. They are allowed to change.