Factory AI's pitch is easy to state and hard to deliver: you describe a task, a Droid does the work, and you review a pull request. That is a different contract from the autocomplete and chat tools most developers adopted between 2023 and 2025. Those tools sit next to you while you drive. A Droid is supposed to drive on its own and hand you the result.
That distinction — pairing versus delegation — is the whole story of autonomous agents in 2026, and it is the right lens for deciding whether Factory belongs in your workflow. We spent time running Droids against real repositories to separate the demo from the day-to-day.
What a Droid actually does
Factory positions itself as an "agent-native" development platform rather than an editor plugin. A Droid accepts a task from where your team already files work — a GitHub issue, a Linear ticket, a Slack message, a Sentry error — then plans, edits across multiple files, runs tests, and opens a pull request you can review like any other contribution.
There are two surfaces. The cloud platform runs Droids asynchronously: you assign work and come back to a PR. The droid CLI runs an agent in your terminal against a local checkout, closer to the interactive loop developers got used to with Claude Code and Codex. The CLI is the better starting point, because you can watch the agent reason and interrupt it before it commits to a bad plan.
What separates a delegation agent from a chat assistant is context gathering. Before touching code, a Droid reads the surrounding files, traces how a function is used, and checks existing tests. On a well-structured repository with clear conventions, that grounding produces changes that match house style instead of generic boilerplate. On a sprawling monorepo with implicit conventions, the same step is where things wander.
The mental model that helps most: treat a Droid like a capable contractor on their first week, not a senior teammate. It knows the language and the tools, but it does not yet know which parts of your codebase are load-bearing and which are safe to touch. Scope the task tightly and the output is strong. Hand it something vague and it fills the gaps with assumptions.
Where the autonomy holds, and where it breaks
The work Droids handle well is the work most teams under-invest in. Dependency bumps with the follow-on code changes. Adding a field through a stack — migration, model, API, type, test. Writing the missing tests for a module that has none. Translating a clear bug report with a stack trace into a fix and a regression test. These are bounded, verifiable tasks where the definition of done is concrete, and that is exactly what an agent needs to stay on track.
The failure modes are just as consistent. Ambiguous requirements are the first. Ask a Droid to "improve performance" and it will pick a metric for you, often the wrong one. Cross-cutting changes that require holding several non-local constraints in mind at once are the second — a Droid will satisfy the constraint it can see and quietly violate the one it cannot. The third is anything where the test suite is weak. Autonomy is only as trustworthy as the verification it runs against; without solid tests, a green checkmark means the code runs, not that it is correct.
The practical consequence is that a Droid does not remove review — it relocates it. You spend less time typing and more time reading diffs critically. For a one-line config change that is a clear win. For a 400-line refactor across six files, the review can cost as much attention as writing it would have, and you carry the added risk that the change looks plausible while being subtly wrong.
Review autonomous pull requests with more suspicion than human ones, not less. An agent's diff is optimized to look correct and to pass the tests it can see — which is precisely the profile of a change that hides its bugs in the gaps your tests miss. Read the parts it did not touch as carefully as the parts it did.
Cost is the other axis to watch. Delegation agents consume far more tokens than interactive ones because they read widely before acting and often retry after a failed test run. A task that feels trivial can still burn meaningful usage when the agent explores a large context or loops on a flaky test. Budget by task complexity, not by how small the change looks.
Should you bring Droids into your workflow?
If your team already files clean tickets, maintains a real test suite, and reviews every change, Droids slot in cleanly as a way to clear the bounded backlog that never gets prioritized. If your tickets are one-line notes and your test coverage is thin, fix that first — the agent will amplify whatever discipline already exists, in both directions.
A reasonable adoption path: start with the CLI on low-stakes, well-tested tasks where you can watch and interrupt. Measure how often you accept a Droid's PR without changes versus how often you rewrite it. That accept rate, tracked over a few weeks, tells you more than any benchmark. Only move to fully asynchronous cloud delegation once you trust the accept rate on a given category of work.
Delegation agents do not replace the interactive, pair-programming style of tool — they sit beside it. Many developers run a Droid for the bounded, ticket-shaped work and keep a fast in-editor assistant for the exploratory coding where they want to stay in the loop the whole time.
The honest summary for 2026: autonomous coding agents have crossed from demo to genuinely useful, but only inside the boundaries you draw for them. The teams getting value from Factory's Droids are not the ones expecting magic — they are the ones who already had clean tickets and good tests, and who treat the agent as leverage on discipline they already practice.
Originally published at pickuma.com. Subscribe to the RSS or follow @pickuma.bsky.social for new reviews.
Top comments (0)