Chat, workflow, agent or team? Picking the right shape for the job.
Most AI disappointment comes from using the wrong shape of tool, not the wrong model. Here's a simple way to pick.
When people tell me AI "didn't work" for them, it's rarely because the model was dumb. It's usually because they used the wrong shape of tool. They tried to run a month-long project in one chat, or built a rigid flowchart for something that needed judgement, or spun up a team of agents to answer one question.
There are really four shapes, and picking the right one is most of the battle. Here's how I think about it.
Shape one: a chat
You ask, it answers. Rewrite this email, explain this clause, brainstorm some names, fix this function. The work starts and finishes while you're sitting there, and you steer every step.
Chats are brilliant and underrated. If the work fits in one sitting and you're happy to drive, nothing beats them for speed. Adding any structure on top would just slow you down.
Shape two: a workflow
A workflow is a process you can draw: when this happens, do that, then this. A new form submission goes into the spreadsheet, gets a reply email and pings the team. You might drop an AI step in the middle (summarise this, classify that), but the route is fixed.
Workflows are predictable, cheap to run and easy to trust, because you can see exactly what will happen. Their weakness is the flip side: they can't cope with anything you didn't draw. The moment a step needs judgement about what to do next, a workflow either breaks or grows a tangle of branches.
Shape three: an agent
An agent gets a goal instead of a route. Research these five competitors. Get this build passing. Draft a proposal from these notes. It decides the steps as it goes, uses tools, looks at what happened and adjusts.
That's what makes agents good at open-ended work, and also what makes them harder to trust. You can't see the route in advance, so you have to judge the result instead, and give the agent clear limits on what it's allowed to do along the way.
Shape four: a team
A team is what you need when the work outgrows one agent's session or one person's attention. It runs for days, involves research and production and review, needs other people's input, or has to keep moving after you close the tab.
The important thing a team adds isn't more agents. It's structure that lasts: someone owns each outcome, goals survive a restart, decisions get recorded, and only the things that need a human reach one. Without that, a "team" of agents is just several chats that don't remember each other.
This is the shape Restless is built for. It's also the most expensive shape, which is why it should be the last one you reach for, not the first.
The common mistakes
Most disappointment comes from one of these four:
- Running a project in one chat. It works for a while, then the thread fills up, summarises itself and quietly forgets your early instructions.
- Drawing a workflow for judgement work. You end up with dozens of branches that still miss the case that actually happens.
- Giving an agent a task a workflow could do. You pay for thinking where a fixed rule would have been cheaper and more reliable.
- Spinning up a team for a single question. More agents means more briefing and more handoffs. For small jobs that's pure overhead.
Shapes change as work grows
Plenty of work starts as a chat and graduates. You brainstorm in a chat, realise it's a real project, and hand it to an agent. The agent finds a step that repeats exactly every time, and that step becomes a workflow. The project grows to involve your designer and a reviewer, and now it's a team.
That's healthy. The skill isn't picking one tool forever. It's noticing when the work has changed shape, and moving it.