Goal Getter started from a simple observation. I was using two different LLMs on the same programming task and manually copying one model's output into the other's context. The results seemed noticeably better than repeatedly prompting a single model. That made me wonder what would happen if the models could cooperate as part of an actual software-development process instead. That became Goal Getter. You open a project, give Goal Getter a vision, and assign one or more AI agents to different stages of the development loop. The basic cycle is: * AI agents propose what should be done next. * They evaluate the proposals and select one. * One of them implements it. * AIs review the result. * Other AIs test it. * The resulting project state goes back into the discussion, and the cycle repeats until the vision is realized. The stages can use different models and different providers. A human can also occupy a role. The vision can be something big or something small, I've tested both with success. If Goal Getter needs a human decision while running, it puts the question into an inbox. The part I've found most interesting is model diversity. For example, I can have one model implement a task while several independent models review or test it. Or I can put models from different providers into the same stage and let them independently do the work. In my own testing, this has often produced better finished software than my previous workflow of repeatedly giving individual tasks to one stronger model. One surprising result was that some crews built around relatively inexpensive models such as GPT-5.6-Luna and DeepSeek-v4-Pro performed very well. In several of my projects, the finished result was better than what I had previously gotten from ordinary one-task-at-a-time prompting with GPT-5.6-Sol, while costing considerably less in inference. These aren't controlled benchmarks, and I don't think cheaper models are somehow universally better. Hard problems still benefit a lot from stronger models. My current hypothesis is that separating planning, implementation, review, testing, and retries can compensate for quite a few weaknesses of any individual model. I originally built a prototype in Common Lisp. Then I eventually moved it to C, and Goal Getter has since been used to work on its own C implementation and to build the web application around it. Other projects I've created using it include: * A browser-based Tetris clone. * A proof-of-work, Hashcash-like, CAPTCHA-like system with a backend and frontend, using WebGPU with a JavaScript fallback. You can check its code on GitHub: https://github.com/Bitwise-LTDA/... * Parts of Goal Getter itself, including the downloadable program and backend. I've deliberately tried to keep the runtime small. The Linux amd64 binary is currently about 20 MB, and the main process typically uses up to 48 MB of RAM, leaving most of the machine available for compilers, browsers, test suites, and whatever the AIs need to run. Goal Getter doesn't include an AI subscription. You bring your own API keys and decide which providers/models should fill each role. There's a free pre-alpha trial. After that, it will cost $10 per machine/month. The same binary can run on a workstation or a Linux server you control. I'm particularly interested in whether other people see the same effect from separating roles and mixing independent models/providers.

Go Getter

Autonomous agentic AI tool

Autonomous agentic AI tool