Skip to content
All projects
Live demoGenerative AIRAGAgentic AIAWS

TrustResponse — Governed Multi-Agent Security Review

A governed multi-agent system that drafts evidence-backed answers to security questionnaires, catches unsupported commitments before they become liabilities, and routes anything risky to a human.

View live demoView on GitHub
TrustResponse — Governed Multi-Agent Security Review interface

Overview

TrustResponse automates the security questionnaire, the step where enterprise deals go quiet for days. It drafts answers from a governed evidence library, runs an adversarial critic over every claim, and escalates anything it cannot support to a person. The point of the product is not how much it answers. It is what it refuses to answer.

It is a live demo running on AWS. The most interesting thing to watch is the agent trace: a retrieval agent, a compliance critic, and a gap agent reasoning out loud, then routing a risky question to a human instead of guessing.

The results below come from a synthetic golden set of 150 questions, not a customer deployment. I frame them as demonstrated behavior.

Business problem

A security questionnaire is a trust document. Every answer is a commitment, and a wrong answer is a future liability. Teams handle them in one of two bad ways: slowly and by hand, or quickly with copy-paste that risks overcommitting. An AI can speed this up, but only if it is more careful than the humans it is helping, not less.

How it works

The pipeline is a governed sequence of agents:

  1. Intake. The questionnaire is parsed into individual questions with category and tier.
  2. Retrieve. A retrieval agent pulls scoped evidence from the Trust Library.
  3. Draft. A drafting agent writes an answer strictly from that evidence.
  4. Critique. A compliance critic tries to break the answer, checking whether the evidence actually supports the claim.
  5. Approve. Supported answers are drafted for review; unsupported or conflicting ones are escalated to the right human owner.

The platform

  • Governed Trust Library. Evidence is scoped by product, region, and tier, with owners and expiry dates, so an answer can never cite something stale or out of scope.
  • Compliance critic. An adversarial agent whose whole job is to reject claims the evidence does not back. This is what catches unsupported commitments.
  • Consistency engine. Keeps answers to related questions from contradicting each other.
  • Human-in-the-loop. Nothing is auto-submitted. The team stays in control of every answer that leaves the building.
  • Template export. Answers export back into the customer's own format, whether that is Excel, Word, or a portal.

Important design decisions

  • The value is in the refusal. A system that drafts 75 percent of answers and clearly flags the other 25 percent is far more useful than one that answers everything and is quietly wrong somewhere.
  • Adversarial critique as a first-class step. A single model asked to "answer carefully" is not as reliable as a separate agent whose only job is to find the hole.
  • Evidence with governance metadata. Owners and expiry turn the trust library from a document dump into something you can actually stand behind.

Evaluation & success metrics

On the synthetic golden set:

  • Roughly 6.2 hours average turnaround, down from about 3 days by hand.
  • About 75 percent of questions auto-drafted with supporting evidence.
  • 47 unsupported commitments caught across a 150-question review.
  • 0 confidently wrong answers, and a retrieval Hit@1 of 1.00 with MRR of 1.00 on the evaluation set.

The evaluation harness, with planted traps the critic is expected to catch, is what makes governance measurable here rather than aspirational.

Architecture

  • A LangGraph-orchestrated agent pipeline runs behind API Gateway on AWS Lambda.
  • Retrieval runs over a managed evidence store; the design favors in-memory retrieval over a standing vector database to keep the demo at roughly zero to a few dollars per month.
  • Delivered as infrastructure as code with AWS CDK and OIDC-based GitHub Actions CI/CD, so there are no long-lived deploy credentials.

Lessons learned

  • An agent system earns trust by what it declines to do. Designing the escalation path was more important than tuning the drafts.
  • Making governance measurable, with an evaluation harness and planted traps, is what turns "we have guardrails" into something you can show a security team.