Agent guide
What is an AI agent?
An AI agent is a goal-directed software loop around a model. It receives context, chooses actions, uses tools, observes results, and keeps working until it reaches a stop condition or needs human review.
The short version
An agent is not just a chat message. Chat asks a model for an answer. An agent gives a model a job, a workspace, tools, constraints, and a way to inspect whether each step worked.
In software work, an agent should be treated like a tool-using collaborator with boundaries, not an unsupervised employee. The review system around it matters as much as the model.
Core parts of an agent
Goal
The outcome the agent is trying to accomplish, such as drafting a plan, fixing a test, or summarizing project knowledge.
Context
The instructions, repository files, docs, conversation, memory, and constraints the agent uses to make decisions.
Tools
Actions the agent can call, such as reading files, editing code, running tests, searching docs, or reporting status.
Control loop
The observe, think, act, verify, and stop pattern that lets an agent move beyond a single response.
Agents in Amistio
Amistio separates coordination from execution. The web app captures wishes, review state, plans, and approvals. The local runner gives the agent controlled access to the paired repository and local tools. That split keeps the product useful without making the cloud execute arbitrary source code.
- Use the web app to define intent and review generated project-brain artifacts.
- Use the local runner to execute approved work inside the repository checkout.
- Use ADRs, plans, and prompts to make the agent's job inspectable before implementation starts.