Learn › How agents work

What is agentic coding, and what is a background agent?

Agentic coding is when an AI system carries out a coding task itself: it reads the code, makes changes, runs things, looks at the result, and continues until the task is done. A background agent does the same work away from your screen and comes back with a finished change. The step after that is an agent that also decides what the task is.

From suggesting to doing

AI help with code has moved through three stages.

StageWhat the AI doesWhat you do
CompletionSuggests the next few linesWrite the code, accept or reject suggestions
Agentic codingCarries out a task you describe, across many filesDescribe the task, review the result
Background agentsCarries out the task without you presentDescribe the task, come back later

What makes something an agent is the loop: it acts, checks what happened, and decides what to do next. A model that only produces text and stops is not one.

What an agent needs

  • Tools. A way to read files, edit them, run commands and run tests.
  • Somewhere safe to work. An agent runs code nobody has reviewed. See what a sandbox is.
  • A way to know it is finished. A passing test, a green build. Without that it stops when it runs out of ideas.
  • Limits. How long it may run, what it may touch, and when it must ask.

Background agents

A background agent takes a task and works on it somewhere else, usually in the cloud. You are not watching each step. You get a pull request at the end.

That changes what matters. With an assistant in your editor, you catch a wrong turn as it happens. With a background agent, you see only the result, so the result has to carry its own proof.

The step after: an agent that finds the task

Every stage above starts with a person describing a task. Someone noticed a problem and wrote a prompt.

A large part of software work has no one to write that prompt. A crash at night. A package that became vulnerable. A build that went red after someone left. The work exists and nobody has asked for it.

Autonomous software maintenance is agents that find those tasks themselves. Bugstack is built this way: it watches your running app and your repository, and when something breaks or goes out of date, it starts work, proves a fix in a sandbox, and opens a pull request.

A coding agentBugstack
Starts whenYou describe a taskSomething breaks or goes out of date
Works onWhatever you ask: features, changes, fixesUpkeep: crashes, failing builds, failed deploys, packages, security
Finishes withA change for you to reviewA pull request with the cause, the fix and a test

You use both. A coding agent builds what you decided to build. Bugstack looks after what has been built, including what the agent wrote.

How much to let an agent do

The useful question is what proves its work, more than how capable it is. An agent whose changes come with a failing test, a clean run and your CI can be given more room than one that reports “done”.

Bugstack makes that a setting per repository: Suggest, Review or Autopilot. Every repository starts on Review, where you merge each change.

Questions

Is agentic coding the same as vibe coding?

They overlap. Vibe coding describes building by describing what you want and accepting what the AI writes. Agentic coding describes how the AI works: in a loop, with tools. You can do agentic coding and still read every line.

Do background agents need supervision?

They need their output checked. The more of that checking is automatic, such as tests and CI, the less of it falls on you.

Is Bugstack a coding agent?

It uses agents: a coordinator, a tester, a fixer and others. The difference is that nobody prompts it. It finds maintenance work on its own and hands back pull requests.

See it work on your own code.

$100 a month

Includes $100 of usage credit every month. Start with a 14-day trial: $50 of credit and a first security audit on us. A card is required.