Bramha is a multi-agentic platform for complex software work. Instead of one model producing a single response, a structured council of agents works the problem across a directed graph, coordinating through a shared, permission-scoped knowledge store.
The Agent Model
Bramha runs 8 executive and 7 worker agents on a DAG conversation model. Executives plan and delegate; workers execute. Crucially, deterministic PA (personal-assistant) agents handle routing and bookkeeping in code rather than through the model — cutting LLM calls per turn from 8 to 0 where a decision doesn’t actually need a model.
The Stack
The platform is built on NestJS and Next.js, with Postgres + pgvector for both relational state and vector retrieval. Row-level security (RLS) scopes every read and write, so agents and tenants only ever touch data they’re allowed to. Orchestration runs on LangGraph, with MCP for tool and context exchange.
Knowledge Pipeline
A managed knowledge pipeline ingests 7 file formats — including CSV, image, PDF, SVG, and ZIP — and 3 external sources: GitHub, SQL, and URL. Everything is browsable and editable inside a visible knowledge-store interface, so what the agents know is never a black box.
Status
Bramha is under active development. Details will sharpen as the public build takes shape.