AutoGPT was the most famous piece of software in the world for roughly six weeks in 2023. The pitch was an agent that pursued a goal autonomously, and the reality was an agent that pursued a goal autonomously into a loop, burning API credits until someone stopped it. It demonstrated something important about what language models could be pointed at and it was not a usable product. Since then the project has been rebuilt into the AutoGPT Platform, a block based visual builder for agent workflows, which is a considerably more sensible thing to be.
The documentation covers the platform architecture, building agents and workflows, the block library and its integrations, the supported model providers, and deployment including a self-hosting guide and a getting started path for the hosted cloud. The core idea is that you compose agents from blocks, each block doing one thing, wired together into a flow. This is a familiar pattern and it works, and it moves the project away from the original unbounded autonomy that made it unusable and towards something with defined behaviour. The visual builder is the main draw.
For someone who wants to connect a trigger to a model call to an API call to a notification without writing code, this is a reasonable way to work, and the documentation walks through it adequately. The block library covers a decent spread of integrations and is growing. The marketplace of existing agents is genuinely useful for learning, because seeing how someone else wired a workflow teaches the pattern faster than any tutorial page. Self-hosting is documented and works if you are comfortable with Docker and a multi service setup.
It is not a one command affair and the guide is honest enough about that. Now the problems, and there are several. Documentation quality is inconsistent. Some sections are current and clear.
Others describe an earlier version of the interface, or reference concepts that have been renamed, or trail off where you most want detail. The project has moved through several significant architectural changes and the written material carries sediment from each of them. You will find yourself checking the repository to work out what is actually true, which is a bad sign for documentation. Licensing is muddled and underexplained.
Most of the repository is MIT. The newer platform components sit under a Polyform Shield licence, which is not open source in the sense most people mean and restricts competing commercial use. This distinction matters if you are building anything on top of it, and it is mentioned rather than explained. Anyone planning commercial use should read the licence itself rather than the documentation's summary of it.
Cost control is the omission I find hardest to forgive, because the original project's single defining failure mode was spending money in a loop. The platform's block structure inherently constrains this better than an unbounded agent did, and the documentation still does very little to teach budgeting, limits, monitoring or what happens when a workflow with a loop runs longer than you expected. Given the history, this should be the most prominent page on the site. Reliability is the other underdocumented area.
Real agent workflows fail. An API times out, a model returns something unparseable, a rate limit hits, a step produces output the next step cannot use. What happens then, how you retry, how you handle partial failure, how you make a long workflow idempotent, all of this is thin. Anyone running these for real work hits these questions in week one.
There is a structural problem with the audience too. The name still carries the fame from 2023, which means beginners arrive expecting the autonomous agent they read about and find a workflow builder with documentation written for someone reasonably technical. The gap between the expectation and the material is wide, and the documentation does nothing to manage it. I should be fair about the direction of travel.
Abandoning the autonomous loop for composable blocks was correct. Building a marketplace was sensible. Keeping self-hosting viable was the right instinct. This is a more honest and more useful project than the one that went viral, and it deserves credit for surviving its own hype cycle, which most projects of that era did not.
My two point nine reflects documentation that is behind its product, licensing that is unclear where clarity matters, and gaps around cost and reliability that are exactly the gaps this project's history should have made it obsessive about closing. If you want a visual agent builder, look at this and also at the alternatives. If you write code, the code based frameworks are better documented, more predictable and easier to debug.