Agents / Trust & SafetyMock demo

WorkflowPilot Safe Agents

A safe-agent demo that turns a business goal into a proposed multi-step workflow, runs only human-approved tool calls, and records every action in an audit trail.

Next.jsTypeScriptClaude tool-useHuman-in-the-loopAudit loggingVercel
Live demoGitHubAll projects

At a glance

ICP
Ops/AI PMs and small teams automating low-risk workflows: drafting emails, creating tickets, summarising notes.
Features
  • Create a workflow goal
  • Agent proposes a tool-call plan
  • Human approval queue (approve / reject / edit)
  • Execute mock tools first (dry-run by default)
  • Full audit trail of proposed → approved → executed
  • Failure / rollback report + safety settings

AI architecture

  1. 1
    Goal
    User states a plain-language objective.
  2. 2
    Planner
    claude-opus-4-8 proposes ordered steps with tool, args, reasoning.
  3. 3
    Approval queue
    No tool runs until a human approves, the core safety gate.
  4. 4
    Execution
    Approved steps run against mock tools (send_email_mock, create_ticket_mock, …).
  5. 5
    Audit log
    Every proposal, approval, execution, and failure is recorded.
  6. 6
    Recovery
    Failure report + rollback suggestions; policy violations are never auto-retried.

Case study

Product problem

Ops and AI PMs want the leverage of agents without the blast radius. The product's job is to make automation trustworthy: every action is previewed, approved, logged, and reversible.

ICP & MVP scope

ICP: a small team automating low-risk workflows. MVP: goal → plan → approval → mock execution → audit. Out of scope (until explicitly approved): real external integrations and unattended execution.

Trust by design

The guardrail metric, unauthorized tool/action attempts, has a target of zero. Trust isn't a feature here; it's the product. The safety settings and audit trail are the things a buyer actually evaluates.

Resume bullets · AI Engineering
  • Built a human-in-the-loop agent with a manual tool-calling loop that gates every action behind explicit approval and an append-only audit log.
  • Implemented a permissions model and mock-tool sandbox so agent safety can be demoed with zero real-world blast radius.
Resume bullets · AI PM
  • Designed a trust-by-design agent product where approval gates, audit trails, and rollback are the core value, not add-ons.
  • Set a zero-tolerance guardrail (unauthorized tool attempts) and framed safety settings as the primary buyer-evaluated surface.