Open Bias enforces agent policies at runtime
Open Bias is an open-source agent alignment proxy that sits between your app and an LLM provider, reads rules from a plain `RULES.md`, and enforces them at runtime. The project is aimed at teams dealing with agent drift, unsafe tool use, and policy violations that prompt-only guardrails cannot reliably prevent. It supports provider-agnostic deployment, live interventions, blocking, and shadowing, with a focus on low-latency enforcement and easy adoption via a single base URL change.
Hot take: this is pointing at the right failure mode. Most “agent safety” stacks observe violations after the fact; Open Bias is trying to make business rules executable at the boundary where the model actually acts.
- –The core value prop is runtime policy enforcement, not better prompting.
- –The `RULES.md` workflow is pragmatic: plain markdown is easier to review, diff, and version than a bespoke policy DSL.
- –The strongest use cases are hard business constraints like discount ceilings, identity checks, and data-leak prevention.
- –The architecture is interesting because it separates enforcement from the model provider, which makes it easier to slot into existing agent stacks.
- –The main question is scope: this looks best for well-specified rules, while fuzzy judgment calls will still depend on evaluator quality.
DISCOVERED
4h ago
2026-04-25
PUBLISHED
7h ago
2026-04-25
RELEVANCE
AUTHOR
Chinmay101202