AutoGPT

You will like this if what you want is automation, not engineering: describe an outcome, or drag blocks together in the Build editor, then run it on a schedule or from a trigger and check the run dashboard in the morning. The daily trait is that it behaves like an automation product, wired to the SaaS tools your company already pays for.

You will not like it if you came for the original agent that wrote and ran code. Today's platform has no terminal, no file edits and no repository awareness. Pick it for ops workflows. Pick goose or Claude Code when the task lives in a repo.

The risk is a borrowed reputation. The 187k stars and 46k forks belong to the original autonomous agent, which the repository now keeps in a classic folder. The platform sold on the landing page is a different product: a block-based workflow builder with no terminal, no file edits, and MCP only as a block you wire yourself. Evaluate it as a new product with a short track record, because that is what it is.

The consequence: do not let the star count stand in for a reference call. What it does right is isolation. Agent runs execute in Docker, so a bad block does not run on your laptop.

The documented architecture has four surfaces. 1. AutoPilot turns a natural-language outcome into an agent. 2. Build is a visual editor where blocks are wired together, so the plan is a graph the user drew rather than one the model inferred. 3. Agents is a dashboard of runs, costs and statuses. 4. Marketplace distributes community agents.

The verification story is thinner. The only benchmark tooling, agbenchmark, sits with the original agent's code, and no published number describes how the platform's agents perform. The design is principled where the user draws the graph; where AutoPilot draws it, the documentation does not say how the result is checked.

This is the rare pivot that kept the brand and swapped the product. The agent that started the category became a workflow platform with Pro at $42.50 a month and Max at $272, both billed annually, plus pay-as-you-go credits for runs. Anchoring at $272 for individuals is a pricing-power test, and annual billing says they want the cash up front.

The moat is the name and the marketplace, neither of which stops Zapier or Make from adding an agent block. Likely acquirer: an automation incumbent buying the brand. Likely pivot: enterprise plans sold through a sales contact. Position: cautious, wait for a revenue signal.

The demo is a chat that builds an automation and runs it on a schedule. Procurement finds that the Team tier, the one with multi-user workspaces, admin roles and seat management, is listed as coming soon. Today that means sixty individual plans, sixty credit wallets metered per run, and an SSO story that exists on the cloud platform as OAuth 2.0 but without SCIM.

CI fit is nil, because this is not a code tool, and the onboarding burden is low for the same reason. Support is email on the standard tier. Not yet. Revisit when the Team tier has a price.

Read the license before the README. The platform directory is Polyform Shield 1.0, which lets me run it for my own use but not sell it as a competing hosted service, and the rest is MIT. That is honest source-available, not open source, and I respect the honesty more than the marketing.

Self-hosting is a shell script and my own keys, and the model list includes Llama, so a local endpoint is in play. MCP is a block on a canvas, not a file I can version, which is backwards. I can fork it for myself; I cannot fork it into a business. That is the deal.