Projects

Agent infrastructure

Voscus

A policy decision between an agent's proposed action and the systems it can change.

Visit VoscusExplore the decision flow

Role
Developer
Built with
Go, Redis, Python, FastAPI
Focus
Pre-execution policy and human approval
Voscus homepage with its pre-execution policy message and a diagram connecting assessment, policy and four decision outcomes.
Voscus's homepage: its pre-execution policy message alongside a diagram of the decision flow.

What I built

Voscus is an agent infrastructure project I built end to end. It combines Go middleware, Python/FastAPI and Redis in a backend policy layer for agent actions.

The problem is concrete: an agent can generate a reasonable-looking request that reaches too many recipients, changes the wrong records or exceeds its authority. The point where a tool call becomes a side effect needs an explicit decision.

The decision boundary

Voscus's public workflow examines the proposed operation, maps its affected scope, and evaluates versioned YAML rules. Its decision path is deterministic, without an LLM choosing the outcome.

  1. Proposed actionA tool call and its parameters reach the policy boundary.
  2. Scope + policyAssess what the action touches and which rules apply.
  3. Decision + reasonReturn an outcome with the context behind it.
  • AllowProceed
  • WarnFlag concerns
  • HoldHuman review
  • StopBlock action
A conceptual view of the published product flow.

The public example holds a large external email send for approval. The reviewer sees the affected audience, matched rule and explanation together.

Product overview and example

Why this boundary matters

This is the kind of backend problem I want to keep working on: make a consequential operation inspectable, define who may authorize it, and give the caller a clear result.

A decision needs stable inputs
For any policy service, repeatability depends on the action, policy version and evaluated context. Changing those inputs can legitimately change the answer.
Approval needs a defined scope
A reviewer needs to know which action they are approving. A changed recipient list or payload raises a new authorization question.
Enforcement needs an integration
The calling system must honor the result. A decision service alone cannot protect a tool invocation that bypasses it.

These are the engineering questions behind the policy boundary. They also connect directly to my interest in platform APIs, reliability and systems that other developers can reason about.

Project scope

This overview combines my end-to-end contribution with Voscus's public product description. The diagram explains that published behavior; it is not a trace from a running deployment. Source-level implementation details, benchmarks and customer results are not published here.

Visit VoscusTalk about the projectAll projects