Open-source LOKĀNTRA lets AI agents act out how people might react to events

What the project is

LOKĀNTRA is an open-source platform, published on GitHub and described in a post on dev.to. You feed it real-world documents or a scenario. It turns them into an interactive knowledge graph, which is a map of the people, groups and ideas in the material and how they connect. It then builds a simulated social environment populated by AI agents.

The idea is to explore how different stakeholders might react to an event, influence one another, and shape where things end up. The developer contrasts this with asking a single large language model (LLM, the kind of AI behind chatbots) for one prediction. Instead, the platform combines several pieces: a multi-agent simulation, graph-based retrieval of information, a system for routing decisions, and reports tied to evidence.

One caution up front: everything here comes from the developer's own write-up. The post lists features but gives no test results, so there is no evidence yet on how well the simulations match real-world outcomes.

From documents to a simulated crowd

The first step is the knowledge graph. According to the post, the developer built a pipeline that generates an ontology, meaning a structured list of the kinds of things and relationships in a document, and converts documents into a graph. It uses TigerGraph, a graph database, and Zep Cloud, a memory service.

Next come the agents. The project uses two existing tools, CAMEL-AI and OASIS, to create simulated personas and model how they interact. The environments imitate Reddit-style and Twitter-style platforms, so the agents post, respond and influence each other in a setting that resembles online discussion.

The platform also keeps graph-backed memory and retrieval, so the agents and the analysis can draw on context from the original material instead of starting cold.

Saving effort with a tiered brain

Running many agents on a large language model is slow and costly. The project's answer is what it calls cognitive swarm routing, a layered way to decide how much thinking each agent needs. The layers are a simple rule-based reflex check, something the post calls LAYA System 1 decision intelligence, and full LLM reasoning. The post does not explain LAYA further, so we will not guess at how it works.

The developer also explored an experimental idea inspired by BrainChip Akida, a brain-inspired chip design. It uses a spiking neural network, a type of AI that mimics how neurons fire, as a reflex gate to cut unnecessary agent computations. The post labels this experimental.

Related cost controls include splitting long documents into representative chunks, managing how much text the model sees at once, generating output in stages, and tracking which decision tier handled each choice. The point of all of this is to use the LLM less often and only where it matters.

Questioning agents and reading reports

Two features speak to the common complaint that AI output is a black box. First, users can interview individual simulated agents and look into the reasoning behind their decisions, opinions and reactions. Second, a ReportAgent gathers evidence from the simulation, writes structured analytical reports, takes follow-up questions, and exports PDFs.

The post is not clear on how faithful an agent's explanation is to what actually drove its behavior. Treat an agent's answer as another output to check, not as proof.

There are also input and integration options. A configurable speech-to-text pipeline from Sarvam AI, called Saaras, lets people describe a scenario by voice. The developer also experimented with viaSocket MCP settings. MCP, the Model Context Protocol, is a common standard for connecting AI tools to other software, and here it supports outside integrations and workflows where a human approves actions.

How much weight to give it

The stack is broad: Python and Flask on the server, Vue.js 3 and D3.js for the interface, Supabase, Docker, and any OpenAI-compatible LLM service. That suggests a working prototype with a real interface, not just a concept. The project also credits TigerGraph's account on social media.

Still, this is a niche project with no independent testing mentioned. Simulated crowds are only as good as the personas, the source documents and the model behind them. A simulation can produce a plausible story without that story being likely.

Where it is useful is as a way of thinking. Exploring possible futures, looking at individual perspectives and tracing how outcomes emerge is a different goal from asking an AI for a single forecast.