What Is a Background Agent?
A background agent is an AI coding agent that works without anyone watching. It starts from an event, a schedule, or a request, runs in its own isolated environment instead of a developer's laptop, and hands back a result for a person to review.
Local Agent vs. Background Agent
Starts from
Local agent: A developer typing a prompt
Background agent: An event, a schedule, or a request
Runs on
Local agent: The developer's laptop
Background agent: Its own isolated environment
Attention
Local agent: Someone watches and steers
Background agent: Nobody watches; a person reviews the result
Scale
Local agent: A few sessions, capped by the laptop and by what one person can watch
Background agent: As many runs as your infrastructure and budget allow
Best for
Local agent: Hard problems that need judgment
Background agent: Repeatable work nobody should have to watch
How It Works
A Trigger Starts It
A pull request, a failed build, an alert, or a schedule starts a run. A person can also delegate a task.
- PR opened
- Build failed
- Alert fired
- Schedule
It Gets Its Own Environment
The agent runs isolated, with the repository, tools, and only the credentials the task needs.
One environment per run
- Agent
- Tools
- Scoped credentials
It Does the Work
It reads the code, makes changes, runs the tests, and tracks what it spent.
- Reads the code
- Makes changes
- Runs the tests
A Person Reviews It
The result comes back as a pull request, a report, or a comment for someone to check.
A person reviews the result
Who Runs Background Agents Today
Stripe · Minions
A run starts from a Slack message in an isolated devbox and ends in a pull request that passes CI.
1,000+
PRs merged each week, all human-reviewed
Ramp · Inspect
Cloud sessions get the same tools an engineer has locally. Separate agents fixed about 100 security issues in 6 days, with people only at PR review.
~30%
of PRs merged to frontend and backend repos
Spotify · Honk
An agent inside Fleet Management writes code changes that roll out across thousands of repositories.
1,500+
merged AI-generated PRs
Uber · Managed agents
Agents review code, fix failing CI, and triage alerts. Uber tracks cost per merged PR and routes work across models to cut cost per session.
9.4x
weekly agent requests across Uber's agentic tools, February to August 2026
What Teams Delegate First
When Something Happens
- Failed build triage
- Incident investigation
- Product checks after a PR
- First-pass code review
On a Schedule
- Dependency updates
- Security fixes
- Flaky test cleanup
When Someone Asks
- Well-defined bug fixes
- Small migrations
Where to Start
Pick One Task
Choose repeatable work that happens every week and affects a small part of the system.
- Dependency update
- Flaky test fix
- Failed build triage
Define Done
Decide what a good result looks like before the first run.
- CI passes
- Tests cover the change
- A person approves
Track Three Numbers
Expand to the next task once these hold for a few weeks.
- Merged without rework
- Share of runs
- Review time
- Minutes per PR
- Cost
- Per task
What to Watch
Cost Grows With Every Run
More runs and parallel tasks increase token spend.
- No markup on tokens
- Cheaper model for simple work
- Cost per task shown
Access Needs a Boundary
An unattended agent can only do what its environment allows.
- Isolated environment
- Scoped credentials
- Self-hosting option
Output Needs a Reviewer
Unattended work still ships under your team's name.
- Pull request or report
- Tests run first
- Human approval
Mistakes to Avoid
Scope
Mistake: Automating everything at once
Instead: One well-scoped task, then the next
Context
Mistake: Vague prompts
Instead: Agent instructions, repo conventions, and good tests
Where it runs
Mistake: A developer's laptop
Instead: An isolated environment for every run
Access
Mistake: Rules the agent is asked to follow
Instead: Boundaries the environment enforces, with scoped credentials
Cost
Mistake: Checking the bill at month end
Instead: Cost per task from day one, with a cheaper model where it's enough
Review
Mistake: Merging agent output unread
Instead: A person approves every change
Build or Buy
Getting started
Build: A few scripts, a CI job, and a local agent
Buy: Connect repositories and pick a trigger
Keeping it current
Build: New models, agent tools, APIs, and security risks every few months
Buy: The vendor keeps the platform current
Who maintains it
Build: Your engineers maintain the platform as well as your product
Buy: The vendor maintains the platform; your team maintains its configuration
Background Agent Platforms Compared
Felan
Models: All major providers, plus local and private models
Where it runs: Managed cloud, or fully self-hosted on Enterprise
Devin
Models: Hosted model catalog; no local models published
Where it runs: Devin Cloud; Outposts keep the agent loop there
OpenHands
Models: Any provider, plus local models
Where it runs: Free Agent Canvas, or paid Teams and Enterprise self-hosting
Tembo
Models: Hosted providers; custom inference endpoints documented for self-hosting
Where it runs: Licensed self-hosted deployment; price not published
Ona
Models: OpenAI models; other providers by exception on Enterprise
Where it runs: Ona-managed VPC in your AWS or GCP account, on Enterprise
Claude Code
Models: Claude models only
Where it runs: Execution only, in beta; Anthropic hosts orchestration
OpenAI Codex
Models: OpenAI models in the cloud; other models in the local CLI only
Where it runs: Only the executor, built against the Agents API; OpenAI runs the agent loop
GitHub Copilot
Models: Multi-model catalog GitHub includes with your plan
Where it runs: Execution only, on self-hosted Actions runners
GitLab Duo
Models: GitLab-managed models to pick from; your own models only if self-hosted
Where it runs: Self-Managed or Dedicated for Government, with self-hosted models
Rovo Dev
Models: Several Atlassian-provided LLMs in the CLI; local models not documented
Where it runs: No published self-hosted runtime
Cursor
Models: Curated hosted models; local models not documented
Where it runs: Self-Hosted Machines run tools only; the agent loop stays in Cursor's cloud
Factory
Models: Multi-model catalog, plus an open-weight pool
Where it runs: Bring your own machine for compute; Factory still coordinates
Coder Agents
Models: Major providers, plus OpenAI-compatible self-hosted endpoints
Where it runs: Self-hosted by default, on every plan including Community
Common Questions
What is a background agent?
A background agent is an AI coding agent that works without anyone watching. It starts from an event, a schedule, or a request, runs in its own isolated environment instead of a developer's laptop, and hands back a result for a person to review.
What are the main background coding agent platforms?
Options include Felan, Devin, OpenHands, Tembo, Ona, Claude Code, OpenAI Codex, GitHub Copilot, GitLab Duo, Rovo Dev, Cursor, Factory, and Coder Agents. They differ in which models you can use, how tokens are billed, and whether the agent can run on your own infrastructure. Felan supports all major model providers, including local and private models. It adds no markup on tokens and runs on its managed cloud or fully self-hosted on Enterprise.
Is a background agent the same as an autonomous coding agent?
Mostly. An "autonomous coding agent" plans and acts on its own. A "background agent" runs unattended in its own environment, started by an event, a schedule, or a request.
Which companies use background agents?
Stripe's Minions produce over 1,000 merged pull requests a week, all reviewed by people. At Ramp, about 30% of pull requests merged to its frontend and backend repos come from its Inspect agent. Spotify's Honk has merged more than 1,500 AI-generated pull requests across its repositories, and Uber runs managed agents for code review, CI fixes, and alert triage.
Do background agents replace Cursor, Claude Code, or Copilot?
No. Developers keep their local agents and copilots for hard problems that need judgment. Background agents take the repeatable work that shouldn't need anyone watching.
What should a team delegate first?
Well-scoped work with a result someone can check: failed build triage, dependency updates, flaky test cleanup, first-pass code review, and well-defined bug fixes.
Are background agents safe to run on production code?
Isolation, credentials scoped to the task, and human review before merging reduce risk, but do not guarantee safety. Check each platform's controls and your own security requirements. Felan's Enterprise plan supports self-hosting for teams that need to keep code on their own infrastructure.
How do you keep background agent costs under control?
Pay provider rates with no markup on tokens, use a cheaper model for simple tasks and a top model for hard ones, run an agent built to use fewer tokens, and track the cost of every task.
Should we build our own background agent platform?
A strong team can build one. The ongoing cost is keeping it current as models, agent tools, APIs, and security risks change every few months, which takes engineers away from your product.
Keep Your Attention for the Hard Problems
Start with one repo and one recurring task.