Project
5D Runtime Policy Engine
Longer, safer, logged autonomous agent loops. Score the action before it runs. Let safe ones through. Review the risky ones. Keep the decision log.
Start with the pain
Approval fatigue and risky tool execution show up together.
Coding agents and autonomous workflows usually fail in one of two ways. They interrupt the user on every harmless action, or they get put into a permissive mode that quietly removes the last useful checkpoint. 5D exists for that exact boundary.
It is the runtime policy layer that sits right before the side effect: score the tool action, return allow or review or deny, log what happened, and optionally hand the risky action to a user or external review agent.
The current runtime goes further than the early public draft. It now includes prompt-injection scanning, output leakage detection, session-level drift tracking, destination policy controls, and a lightweight @gate decorator that can wrap Python functions directly before execution.
Best fit
- Coding-agent and ops-agent power users tired of approval spam and blind trust
- Agent builders who need one policy layer across tools and frameworks
- Teams productionizing autonomous workflows with real side effects
Why teams care
- Fewer useless approval prompts on harmless reads and searches
- Safer shell, write, and network actions without blindly enabling auto mode
- Longer autonomous loops with a real log of what the agent tried to do
What it does
Policy at the action boundary
Use 5D when your agent can write files, run shell, call external APIs, or touch sensitive tools.
- Let low-risk actions run without burning human attention
- Route risky actions to a user inbox or an external review agent
- Block obvious foot-guns before the tool executes
- Keep a decision log for every gated action
- Track session drift before the runtime normalizes bad patterns
- Scan for prompt injection and output leakage around the tool boundary
Why it is different
Local, configurable, and not locked to one provider
- Runs locally and stays configurable
- Not tied to one provider or one hosted approval mode
- Can hand risky decisions to a user or external review agent
- Scales across multiple agents with one runtime policy shape
- Validated through a growing offline benchmark harness and runtime test suite
Available integrations
Use the runtime you already have
Open source and disclaimer
Licensed openly. Operated at your own responsibility.
5D is open source under Apache-2.0 and provided as-is. There is no warranty that it will prevent incidents or make a runtime safe by itself. The operator remains responsible for review, testing, configuration, sandboxing, and deployment.
Try 5D
Put runtime policy where the side effect begins.
5D is built to be local, configurable, provider-neutral, multi-agent ready, and able to hand risky actions to a user or external review agent. That is what makes longer loops more usable without turning them into a blind trust exercise.