The project involved migrating the Opencode terminal agent from Go to Emacs Lisp to create a local, autonomous development assistant. The development took place from late April to late October 2026, resulting in a prototype consisting of approximately 22,000 lines of code.
The system utilizes Apache Jena Fuseki with Lucene and HNSW extensions as a triplestore for managing sessions and project artifacts. This infrastructure supports SPARQL queries, enabling potential information exchange between agents in a distributed network.
The agent operates as a finite state machine, with processes defined by OWL/RDF/TTL ontologies. The workflow includes task analysis, code generation, requirement specification, and automated testing cycles, with user reviews integrated into the process.
For local execution, the system relies on an OpenAI-compatible API via Ollama. The ministral-3:8b model was identified as the most effective for generating Lisp code, achieving speeds of 40–50 tokens per second on an RTX3060 12Gb GPU.
After approximately 500 iterations and 2,000 completed tasks, the prototype remains unable to modify its own codebase. However, the established infrastructure allows for ongoing development and testing of autonomous software engineering workflows.
