Agent Scout

The Pre-Action CI/CD Firewall Gate & In-Flight Diff-Repair Middleware for Autonomous Agents on SharedNet.

Comment

AgentScout: The Pre-Action CI/CD Firewall Gate for AI Agents

“Challenge it before it acts.”

AgentScout is an autonomous pre-action verification and in-flight diff-repair middleware for AI agents.

It intercepts agent drafts before external execution, audits claims against multi-source ground truth, detects contradictions and hallucinations, blocks unsafe outputs, and surgically repairs verified errors before the response is shipped.

The Core Problem in Agent Economies

When autonomous agents interact and transact across SharedNet / SharedOS, a single hallucinated specification, outdated price, or fabricated capability can:

  1. Drain limited Arena credits through incorrect decisions or broken seller claims.
  2. Trigger unauthorized or unsafe tool actions that violate environmental constraints.
  3. Cause cascading errors across multi-agent pipelines.

Existing verification systems primarily focus on determining whether a claim is trustworthy.

AgentScout goes one step earlier.

Instead of allowing an unverified output to reach the user, another agent, or an external tool, AgentScout acts as a pre-action firewall that verifies and, when possible, repairs the output before it ships.

What AgentScout Does

                    AI AGENT
                       |
                       | Draft / Proposed Action
                       v
              +--------------------+
              |    AGENTSCOUT      |
              |   FIREWALL GATE    |
              +---------+----------+
                        |
            +-----------v-----------+
            | 1. Atomic Claim       |
            |    Deconstruction      |
            +------------------------+
            | 2. Ground Truth        |
            |    Retrieval           |
            +------------------------+
            | 3. Multi-Perspective   |
            |    Verification        |
            +------------------------+
            | 4. Contradiction &     |
            |    Boundary Detection  |
            +-----------+------------+
                        |
                +-------+-------+
                |               |
              VALID        CONTRADICTED
                |               |
                v               v
          CLEARED FOR        BLOCKED
           SHIPMENT             |
                                v
                       IN-FLIGHT REPAIR
                                |
                                v
                         RE-VERIFICATION
                                |
                                v
                            APPROVED

1. Pre-Action CI/CD Firewall Gate

POST /firewall/gate

AgentScout evaluates a candidate output against configurable reliability and safety thresholds, with a default clearance threshold of 80/100.

Contradicted propositions are intercepted before they can be delivered or used for downstream execution.

Possible outcomes include:

  • APPROVED_CLEAN
  • BLOCKED_UNSAFE
  • REPAIRED_AND_APPROVED

This turns verification into a deployment gate for AI-generated outputs.

2. Surgical In-Flight Diff Repair

POST /repair

Detection is only half the problem.

When AgentScout finds a contradiction, it attempts to repair the affected portion of the response using verified ground truth.

Example

An agent proposes:

“Sony WH-1000XM5 has 50 hours of battery life with ANC enabled.”

AgentScout discovers the verified specification:

30 hours with ANC enabled

Instead of simply returning FAILED, AgentScout performs an in-flight correction:

- 50 hours with ANC enabled
+ 30 hours with ANC enabled

The repaired output is then re-verified before clearance.

3. Adversarial Arena Stress Testing

POST /attack

AgentScout can actively challenge an agent's output rather than simply accepting supporting evidence.

The adversarial verification layer hunts for:

  • Numerical boundary violations
  • Contradictory specifications
  • Deprecated or outdated information
  • Unsupported claims
  • Ungrounded marketing superlatives
  • Temporal inconsistencies

The goal is simple:

Try to break the answer before the real world does.

4. Ed25519 Cryptographic Clearance Dockets

Every verified response can be accompanied by a cryptographically verifiable clearance record.

AgentScout combines:

  • Ed25519 asymmetric signatures
  • SHA-256 hash-chain provenance
  • Audit identifiers
  • Verification results
  • Evidence references

Downstream agents can retrieve the public key through:

GET /api/v1/public-key

This creates a portable proof that the output passed AgentScout's verification pipeline.

Deep SharedOS Native Integration

AgentScout is designed around the SharedOS security model.

Deny-by-Default Capability Authorization

Internal tool calls and resource access are authorized through the SharedOS capability model against durable capability grants.

Sandboxed files Resource Plane

Evidence, benchmarks, and audit resources operate inside controlled resource boundaries designed to prevent path traversal and unauthorized filesystem access.

Fail-Closed Authorization

If the authorization datastore becomes unavailable, AgentScout fails closed and returns an authority_unavailable refusal rather than silently continuing without authorization.

Durable Audit Sync

Decision records are persisted locally in:

.sharedos/audit_log.jsonl

and can be synchronized with the SharedOS audit service using bounded timeouts, retries, and preserved event IDs.

Agent Access & Transport

AgentScout is accessible through multiple agent-facing interfaces.

1. SharedNet Arena Room Watcher

AgentScout can operate directly inside a SharedNet Arena room through the SharedNet watcher:

npx -y sharednet@latest watch \
  --on message \
  --run "python agentscout_arena_watcher.py" \
  --reply

The watcher supports commands such as:

@agentscout gate <draft>
@agentscout trial <claim>

as well as structured service requests.

2. Model Context Protocol (MCP)

AgentScout exposes MCP interfaces for agent-native integration.

HTTP

/mcp

Stdio

python mcp_server.py

This allows compatible AI agents to invoke AgentScout verification capabilities directly as tools.

3. Machine-Readable Discovery

AgentScout exposes machine-readable service metadata:

Interface Endpoint
Agent Card GET /.well-known/agent.json
Service Catalog GET /api/v1/listing
Public Key GET /api/v1/public-key

This allows autonomous agents to discover AgentScout's capabilities without relying on a human-readable webpage.

4. Zero-Config CLI

AgentScout can also be invoked directly from the command line:

python agentscout_cli.py gate \
  "The Eiffel Tower was built in 1950."

Verification & Test Metrics

AgentScout has been tested across API endpoints, adversarial black-box scenarios, committee verification, and SharedOS capability authorization.

Current Metrics

  • 52 / 52 automated tests passing
  • 100% green test suite
  • Median response latency: <1.5 seconds
  • Zero external API key requirement
  • Deterministic offline knowledge-corpus fallback
  • Fail-safe behavior when evidence retrieval is unavailable

The system is designed to degrade safely rather than silently approve unverifiable claims.

Why AgentScout?

AI agents are increasingly capable of making decisions and executing actions autonomously.

But there is a fundamental problem:

An agent should not be the only system responsible for deciding whether its own output is trustworthy.

AgentScout introduces an independent verification layer between:

Agent Thought
     |
     v
AgentScout
     |
     v
Verified / Repaired Output
     |
     v
External World

Instead of asking an agent:

“Are you sure?”

AgentScout asks:

“Can your claim survive an adversarial verification process?”

AgentScout vs. Traditional Verification

Traditional verification:

Claim
  |
  v
Check
  |
  v
PASS / FAIL

AgentScout:

Claim
  |
  v
Evidence
  |
  v
Adversarial Challenge
  |
  v
Contradiction Detection
  |
  v
Surgical Repair
  |
  v
Re-verification
  |
  v
Cryptographic Clearance

It doesn't just identify the failure. It attempts to fix it before the failure reaches the real world.

Built For the SharedOS / MentorMates Arena Hackathon

AgentScout is built as an agent-native verification service for autonomous systems operating within the SharedOS / SharedNet ecosystem.

Don't ask an AI agent to trust itself.

Challenge it before it acts.

Project Links

Machine-Readable Agent Card

/.well-known/agent.json

Public API Listing

/api/v1/listing

Public Key

/api/v1/public-key

Repository

https://github.com/ShaikhWarsi/Shared-Scout

Discord username - axinto

Built With

  • ai-agents
  • ci/cd
  • cryptography
  • ed25519
  • fastapi
  • json-rpc
  • mcp
  • node.js
  • pytest
  • python
  • rest-api
  • security
  • sharednet
  • sharedos
  • tailwindcss
  • typescript
Share this project:

Updates

Submission history