Inspiration
AI agents are increasingly able to make decisions, spend resources, and call other agents on behalf of users. But an agent can become a judge of its own reasoning: the same system proposing an action may also decide that the action is safe.
We wanted to build an independent layer that asks a different question:
Before your agent acts, can another agent try to prove it wrong?
That idea became BLACKBOX — Trust Before You Act.
What it does
BLACKBOX is an adversarial decision-auditing service for AI agents.
A calling agent provides:
- a proposed decision
- its reasoning
- important constraints
BLACKBOX then independently:
- Identifies the decision domain and risks
- Generates targeted evidence searches
- Retrieves and filters relevant evidence
- Looks for evidence supporting and contradicting the decision
- Identifies decision-specific blind spots
- Calculates a deterministic risk score
- Returns GO, NO-GO, or UNCERTAIN
The goal is not to make the decision for the calling agent. The goal is to make the agent's decision harder to make recklessly.
How we built it
BLACKBOX is implemented as a callable service with a FastAPI backend and a SharedOS-compatible agent interface.
The core pipeline is:
Decision → Domain Extraction → Targeted Retrieval → Relevance Filtering → Evidence Analysis → Risk Assessment → Verdict
A major engineering challenge was preventing generic words from polluting evidence retrieval. For example, an AI coding-security decision could accidentally retrieve articles about "medical coding" or "service credits." We solved this with domain-first query generation and domain-aware relevance filtering.
BLACKBOX also fails closed: when relevant independent evidence cannot be established, it returns UNCERTAIN rather than inventing evidence.
SharedOS / SharedNet
BLACKBOX is designed for the agent-to-agent economy.
SharedNet provides the discovery and communication layer, while SharedOS provides the permission-controlled execution boundary and audit trail.
Our service is:
- Service: BLACKBOX AUDIT
- Price: 8 Arena credits
- Purpose:
blackbox-decision-audit - Agent:
agent://blackbox.audit
This creates a simple economic use case: an agent can spend a small number of credits on BLACKBOX before spending significantly more credits on another service or taking a risky action.
Example
We tested BLACKBOX with an agent considering spending 40 Arena credits on an AI coding service claiming it could automatically fix repository-wide security vulnerabilities.
BLACKBOX returned an UNCERTAIN assessment and identified risks including:
- security regressions
- false negatives
- repository-wide blast radius
- dependency changes
- independent verification
- rollback requirements
Instead of blindly approving the purchase, BLACKBOX recommended testing the service on a non-production repository, independently scanning the results, running regression tests, and maintaining a rollback path.
What we learned
The biggest lesson was that an agentic audit system is only as good as the evidence it considers. Retrieval cannot simply search for words from a decision. It has to understand the decision's domain and reject superficially similar but irrelevant information.
We also learned that uncertainty is a useful result. A trustworthy agent should be willing to say "I don't know yet" rather than manufacture confidence.
What's next
BLACKBOX can become a reusable due-diligence layer for agent economies: evaluating purchases, tool calls, deployments, security actions, recommendations, and other consequential decisions before they happen.
BLACKBOX doesn't make the decision for your agent. It makes the agent earn the right to make it.
Log in or sign up for Devpost to join the conversation.