FIELD NOTES / BUILDING GOOD AGENTS

An agent is still
a model + tools + a loop.

The difference between a thin wrapper and a serious agent is the system built around that loop.

The model chooses what to do next. The surrounding software gives it useful capabilities, keeps the task alive, controls what may happen, and checks the real result.

SAME MODEL · DIFFERENT SYSTEM

A tool call is not yet a serious agent.

×
LIMITED FOR OPEN-ENDED WORK

One-shot tool caller

User
Model
One tool
Answer
STOPThe first result becomes the answer.
  • No feedback loop
  • No durable state
  • No recovery
  • No proof of outcome
✓
TARGET PATTERN

Goal-driven agent

GoalAgent loop
Choose
Act
Observe
Update
Verify
FINISH ONLY WHENGoal proved—or honestly blocked.
  • Adaptive path
  • Saved progress
  • Safe recovery
  • Verified outcome
FROM ARCHITECTURE TO CODE

Build it. See what each piece adds.

Choose a layer, make a decision, and inspect the files it creates.

Loading saved design…
02 / UNDERSTAND & CHOOSE

What does a good result look like?

Write the job you want done. Include what the agent must produce and what it must leave alone. A goal is useful when someone can check whether it was achieved.

This becomes goal.md. The loop reads it before choosing what to do.

Add this layer to create its files →
03 / YOUR PROJECT0 files
▱

Your project starts here.

Add the business goal to create goal.md. Each new layer adds its files to this same project.

agent-project/
└─ waiting for your first layer

Illustrative code. Adapter names show where your framework and services connect.

THE SMALLEST SERIOUS BUILD

Start with one agent.

Strong model+Agent loop+Useful tools+Saved state+Access checks+Result checks

Then add a code workspace, skills, deeper context handling, or another agent only when the task proves it needs them.

Why these patterns are credible

These patterns are a vendor-neutral synthesis of published agent loops, tool design, context management, durable execution, and outcome evaluation. They are engineering starting points, not proof that a particular agent is reliable.

This page explains a design. It does not run or certify a live agent.