Human intent
What outcome is actually needed?
My AI operating view
AI is a worker inside a system. Not the whole system.
The useful question is not “Can the model answer?” It is “Can the complete workflow produce the intended result, safely, repeatedly and with evidence?”
My mental model
Every useful AI workflow crosses clear boundaries.
The agent should understand enough to choose. It should not need to know every implementation detail of the system doing the real work.
Agent decision
Interpret, plan and choose an action.
Tool boundary
Validate and translate the request.
Existing service
Perform the actual business operation.
Evidence
Prove what changed and what did not.
Owner closes
Accept, recover, escalate or stop.
My definition of done
The user’s intended outcome happened correctly—not merely that the model answered, a tool was called, or an API returned success.
Division of work
Use AI for judgment. Use code for guarantees.
Important controls should not depend on whether a prompt is followed perfectly on every run.
Interpret and adapt
- Understand natural-language intent
- Plan within the current task
- Select an appropriate tool
- Adapt when new evidence appears
- Explain the result to a person
Enforce and protect
- Validate required information
- Apply permission boundaries
- Limit retries, time and cost
- Protect irreversible actions
- Record traceable system evidence
Define and accept
- Define business meaning
- Set acceptable risk
- Supply domain knowledge
- Resolve ambiguous cases
- Own the final outcome
How I apply it
Start from work closure, then design backward.
I do not start by choosing a model. I start by identifying the work, evidence, risk and ownership around the model.
Define the outcome
Write what “done” means in system and user terms.
Bound the task
Give the agent one clear objective and only the tools it needs.
Contract every handoff
Define required input, possible output and failure meaning.
Require independent proof
Check system state rather than trusting the agent’s own confidence.
Name the owner
Decide who acts when each boundary rejects or fails.
Measure repeated results
Compare quality, rework, time, safety and cost across many runs.
How I release it
Autonomy is earned through evidence.
Responsibility remains with people until the workflow proves it can safely take on more authority.
Assist
AI provides a recommendation. A person performs the action.
Prove usefulnessPrepare
AI creates a plan or change. A person approves execution.
Prove plan qualityExecute
AI performs one bounded task. A person reviews the evidence.
Prove repeatabilityAutomate
Only stable, reversible and observable tasks run automatically.
Prove safe operationFailure without blame games
Make every failure attributable.
Middleware should faithfully carry and explain a request. It should not become the automatic owner of every failure around it.
Wrong choice or arguments
Improve instructions, tool description, examples and evaluation cases.
Invalid translation or contract
Fix validation, mapping, failure types and traceability.
Execution rejected or unavailable
Preserve the real failure and route it to the service owner.
Correct call, wrong user outcome
Revisit the workflow and definition of done—not only the prompt.
What I do not accept as proof
A confident answer is not production evidence.
The most dangerous failures are often workflows where every technical component reports success while the user receives the wrong result.
× One successful demonstration
× A tool call treated as completion
× AI acting as its only evaluator
× A generic “cannot call” error
× Ownership decided after failure
× Adoption measured without outcome
The simple version
Let AI reason. Make the system prove.
My role is to connect those two parts: enough AI freedom to handle real work, and enough engineering discipline to make the result safe, visible and owned.