AI assistance vs autonomous project management

The line in AI assistance vs autonomous project management is about who makes the next decision: you, or software acting on your behalf. With assistance, an agent writes code inside a task you chose, and you decide what happens next. With autonomous project management, software would set priorities, break goals into tasks, assign agents and certify that work is done. Between those ends there are several useful levels, and most confusion comes from describing a lower level with the vocabulary of the highest one. This page lays out five levels, what each one needs before you can trust it and what has actually been shown at each level, including in BRAIN FLUID's own development work. The aim is a vocabulary you can use to judge any tool or workflow, including your own scripts. For the related question of which category of tool you need, see coding agent versus project manager.

Five levels of delegation

Level 1, assisted edits. The model suggests code or answers questions; you apply and run everything. Every decision is yours.

Level 2, delegated tasks. You start an agent session on one task, such as 'add a VAT line to the invoice PDF'. The agent edits files and runs commands within that session; you review the result and decide whether it is done. This is how most people use Claude Code and Codex CLI.

Level 3, card-level automation. The agent also updates the task record: it marks the task started, reports progress and blockers, claims completion and records follow-up work it discovered. The record now changes without you, but you still choose which task runs.

Level 4, bounded coordination. A runner works through an explicit list of tasks you approved, in dependency order, starting each only after the previous one has a verified completion, within limits on launches, retries, time and failures. Choice of next task moves to software, within your list.

Level 5, autonomous project management. Software decides what matters, decomposes goals, selects agents and certifies done. Your role shrinks to setting goals.

What each level must prove

Each level moves one decision from a person to software, so each needs evidence that the software makes that decision safely.

Level 2 needs a declared scope and a review you actually do. The risk is drift beyond the task; see keeping agents within scope.

Level 3 needs independent checks, because the agent's status reports are assertions. An agent that says 'done, tests pass' is right most of the time, and the times it is wrong are the expensive ones. A completion should only count when the checks ran outside the agent's control.

Level 4 needs everything from Level 3 plus an explicit task list, enforced limits, clear stop conditions and recovery from interruption. The runner must refuse to pick up tasks nobody approved, including tasks discovered during the run.

Level 5 needs things that remain open problems: certifying that arbitrary work is correct needs checks at least as good as the work, and prioritization needs value judgments about your users and your business. A tool claiming Level 5 should be able to show you how it decides both, on a real project, with failures reported.

What has and has not been shown

Levels 1 and 2 are everyday practice. Their limits are well known: agents misread tasks, drift in scope and overstate success, and good teams compensate with review.

Levels 3 and 4 are where demonstrations need the closest reading. Ask whether the demo used a live model or a scripted stand-in, how many tasks it covered, how complex they were and whether completion was checked independently. BRAIN FLUID's own development work is a useful example of honest reporting at these levels: a small native rehearsal where each CLI completed one arithmetic task through a card workflow, and a three-card coordinated run driven by a scripted agent. Neither is evidence of complex autonomous implementation, and they are not described as such.

For Level 5, apply the questions from the previous section. BRAIN FLUID does not attempt it: no global prioritization, agent selection, decomposition or objective certification. If any product shows you Level 5, ask to see the failures and how the system decided a task was done. A demo that only shows successes tells you very little. The page on evaluating AI coding workflow tools has a broader checklist.

Choosing the right level for your project

Pick the level by the strength of your checks, not by ambition. Work at Level 2 for anything where 'correct' is hard to test: new features with fuzzy requirements, UI changes, design decisions. Most project work belongs here, and doing it with good records (a task per change, scope, recorded outcomes) is a large improvement on chat-only work.

Move a category of work to Level 3 when its checks are reliable and cover what you care about, so a verified completion actually means something. Dependency updates with a strong test suite, or mechanical refactors with type checking, are reasonable candidates.

Try Level 4 only for a short chain of well-specified tasks with independent checks between them, a limit on everything and a run you can read in full the next morning. Treat the first several runs as experiments. Compare interactive and unattended sessions before deciding.

Whatever level you choose, keep the final acceptance decision human. Pairing styles are covered in AI pair programming versus delegation.

QUESTIONS

Can AI agents manage a software project on their own?

Not in a way you should rely on today. Agents handle individual tasks well enough to be useful, and runners can work through an approved list with checks between tasks. Deciding priorities, breaking goals into tasks and certifying that arbitrary work is correct remain open problems. Keep those decisions with a person and let agents work within them.

What is the difference between an AI agent and an autonomous agent?

Most coding agents act autonomously inside one session: they choose which files to read, edit and test. What they do not do on their own is choose the next task, approve their own scope or accept their own work. Autonomy is a question of which decisions move from you to the software, so it helps to name them one by one.

How should I judge a demo of an autonomous coding agent?

Ask whether it used a live model or a scripted stand-in, how many tasks it covered and how hard they were, who decided the tasks were done and with what checks, and what failed. A demo that shows only successes, or where the agent certifies its own work, tells you little about how the system behaves on your project.

RELATED GUIDES

• What is a project management harness?
• Cards vs scheduled tasks: what starts an agent?

RELATED HARNESS PAGES

• AI coding agent vs AI project manager AI coding agent vs AI project manager: an agent does one task, a project layer holds plan, context, checks and record across many. Neither replaces the other.
• Bounded autonomy: limits every unattended agent run needs Bounded autonomy: limits every unattended agent run needs, from an explicit task list to launch, retry, time, failure and spend limits.
• Running dependent AI tasks in sequence Running dependent AI tasks in sequence: start the next task only after a verified completion, one assignment at a time, with an explicit policy for blocks.
• AI pair programming vs delegating tasks to agents AI pair programming vs delegating tasks to agents: what each mode needs from you, where each pays off, and the signals that tell you to switch.
• How to evaluate an AI coding workflow tool How to evaluate an AI coding workflow tool without trusting demos: matched pairs, independent checks, every failure kept, and limits stated plainly.