FARAN

FARAN

Inspiration

What it does

How we built it

Challenges we ran into

Inspiration

AI assistants are useful in the moment, but most of their work disappears after the answer. People still have to translate goals into plans, reconnect prior knowledge, track execution, and preserve what was learned. FARAN was created to make that entire cycle one coherent workspace instead of another chat window.

What it does

FARAN is an AI Second Brain that turns a natural-language goal into a durable multi-agent workflow. A Planner frames the outcome, Research decides when evidence is needed, Memory retrieves relevant long-term context, Task creates an ordered execution path, Reasoning validates the result, and Writer returns one useful response.

Plans, tasks, evidence, workflow state, and results remain available for future semantic retrieval. FARAN also supports idea connections, retryable queued workflows, schedules, conversation continuity, and evaluation from expert corrections.

How we built it

FARAN uses FastAPI with clean route, service, repository, agent, and tool boundaries. GPT-5.6 Sol coordinates quality-critical workflows through the OpenAI Agents SDK and Responses API. GPT-5.6 Luna handles bounded, schema-validated analysis.

Agent handoffs, strict Pydantic outputs, function tools, persisted response IDs, prompt caching, reasoning controls, and context compaction form the AI runtime. SQLite, SQLAlchemy, and Alembic provide durable storage. The memory architecture includes embeddings, vector records, semantic retrieval, and automatically discovered Idea Connections.

Codex was our primary engineering partner. It audited and refactored the architecture, implemented the agent and memory layers, integrated GPT-5.6, diagnosed live workflows, built the workspace interface, and repeatedly verified the complete test suite.

Challenges

The hardest part was making agentic behavior durable without creating an uncontrolled chain of prompts. Every agent needed a clear responsibility, typed contract, safe tool boundary, and recoverable execution state while still returning one coherent result.

Accomplishments

Real GPT-5.6 multi-agent orchestration
Long-term memory and semantic retrieval
Automatically discovered Idea Connections
Durable queue, retry, recovery, and scheduling
Correction-to-regression evaluation
Bilingual Second Brain workspace
73 automated tests and passing GitHub CI

What we learned

Agent count is not intelligence. Useful autonomy comes from narrow responsibilities, typed contracts, explicit tools, durable state, and evaluation. Memory must also be more than a transcript: it should be classified, retrievable, connected, and useful to the next decision.

What's next

FARAN will expand toward permissioned workspace actions, stronger retrieval evaluation, multi-user storage, and approval-aware integrations for files, calendars, and productivity tools.