Mari blog

We need background agents

Daniel Henneberger · 2026-05-07

For a year now, the conversation around AI in software has been dominated by the assistant: a chat window that helps you while you're already at the keyboard. That's useful. It is also, on its own, a profoundly limited mental model.

The next leverage point isn't a smarter chat window. It's background agents — patient, scoped, observable workers that run when nobody is watching, and that ship pull requests by morning.

What we mean by "background agent"

A background agent is software that:

1. Runs on its own clock, not on yours.
2. Owns a narrow contract — a specific outcome, not a general-purpose assistant.
3. Produces reviewable artifacts — a PR, a ticket update, a Slack draft — never silent state changes.
4. Stops on uncertainty rather than hallucinating forward.

Read those again. Each one is a refusal of how today's tools work.

Foreground tools are interruptive by design. Background agents are the opposite — they are quiet until they have something worth showing you.

Why now

Three things changed in the last 18 months:

• Long-context models that can hold a non-trivial slice of a repo and its history at once.
• Tool use APIs that make it routine to chain reads, writes, and shell calls.
• OIDC + scoped credentials that let an agent act on infra without a human handing it long-lived secrets.

Put those together and the agent is no longer a research demo. It's a deployable unit. A junior engineer with no working memory but unlimited patience.

Where Mari fits

Mari's whole bet is that documentation is a background-agent-shaped problem. Consider:

// today: every team
const docsAreStale = stalenessOf(docs) > 30 // days
const onCallLearnsTheHardWay = true

The work is repetitive, low-glamour, and — critically — bounded. There's no ambiguity about what "the runbook for service X is wrong" means. There's a ground truth in the code, the PRs, and the Slack threads. An agent can find the gap, draft the patch, open the PR, and get out of the way.

That's a job description.

What we still need to figure out

Background agents demand a different operational model:

• Cost telemetry: agents consume tokens; teams need budgets the way they have AWS budgets.
• Failure modes: silent drift is worse than a noisy crash. Every agent must emit a heartbeat.
• Authority: which agents can merge? Which can only suggest? This needs to live in code, reviewed like any other policy.
• Provenance: when a doc says X, the reader must be able to trace it back to the source the agent used.

The last two stopped being roadmap and became the product. Mari keeps a record of everything your AI writes: every fact gets a source, an owner, and a sign-off, and fixes land as PRs that wait for your review. Authority is the sign-off. Provenance is the record. The first two we'll keep writing about.

If you'd rather see it live, book a demo, or open an issue on GitHub and pull on the thread.