Future AGI is an open-source platform for testing, monitoring, and improving AI agents. It puts simulation, evaluation, guardrails, tracing, an LLM gateway, and prompt optimization on one platform, so that traces from live traffic feed back into evaluations and the next version of an agent. The lifecycle it covers runs simulate, evaluate, protect, monitor, and optimize.
It runs as a managed cloud service or self-hosted, either as a single app container next to Postgres and ClickHouse or as one container per service, with a Helm chart for Kubernetes and an air-gapped mode. Its interfaces are open: OpenTelemetry OTLP for traces, OpenAI-compatible HTTP for the gateway, and Postgres and ClickHouse SQL for storage, so any layer can be swapped for your own stack.
The platform is a Django and Channels backend with a React front end, a Go gateway, and a Go trace collector, with Redis for live state and Temporal for jobs. Separately packaged SDKs cover tracing (traceAI, for Python, TypeScript, Java, and C#), local testing and red teaming (agent-learning-kit), the platform API, and the gateway client.
| Capability | Future AGI | Langfuse | Phoenix | Braintrust | Helicone |
|---|---|---|---|---|---|
| Agent simulation | Yes | No | No | No | No |
| Voice agent eval | Yes | No | Cookbook | No | No |
| LLM gateway built in | Yes | No | No | Yes | Yes |
| Guardrails built in | Yes | No | No | No | No |
| Prompt optimization | Yes | No | No | No | No |
| Self-host | Yes | Yes | Yes | No | Yes |
Features
- Simulation: runs multi-turn conversations against personas, adversarial inputs, and edge cases, for text agents and voice agents on LiveKit, VAPI, Retell, and Pipecat
- Evaluation: more than 50 metrics behind one evaluate() call, including groundedness, hallucination, tool-use correctness, PII, tone, and custom rubrics, scored by LLM-as-judge, heuristics, or ML models
- Guardrails: 18 built-in scanners for PII, jailbreaks, and injection, plus 15 vendor adapters such as Lakera, Presidio, and Llama Guard, inline in the gateway or as a standalone SDK
- Tracing: OpenTelemetry-native instrumentation for more than 50 frameworks, with span graphs, latency, token cost, and live dashboards
- Agent Command Center: an OpenAI-compatible gateway for more than 100 providers with 15 routing strategies, semantic caching, virtual keys, MCP, and A2A
- Prompt optimization: six algorithms (GEPA, PromptWizard, ProTeGi, Bayesian, Meta-Prompt, and Random) that use production traces as training data
- Deployment options: Docker Compose in standalone or distributed mode, Kubernetes through Helm, VMs on AWS, GCP, or Azure, and air-gapped installs
- Platform access: annotation queues, dashboards, an agent playground, dataset experiments, and access to the platform over MCP
