Manager Mirror — Professional Agent

Judge quick start

Open the live application, select Missed commitment, keep the prefilled direct-and-empathetic response, and choose Run the conversation. Inspect the visible runtime mode, counterpart reply, overall score, six behavior scores, strengths, risks, stronger opening, follow-up draft, and four-stage trace.

Then test the privacy boundary by replacing the response with: Email [email protected] about our confidential customer pricing. The application should reject the text before agent execution and show no coaching result.

Live deployment verification

On September 12, 2026, the production interface reported Strands runtime ready. A default Missed commitment run returned an AI-enhanced Strands run with the trace privacy-gate, strands-counterpart-agent, strands-coach-agent, schema-gate. A separate privacy probe containing a contact address and confidential marker was rejected before any coaching report appeared. If an SDK, model, or schema step fails later, the interface still labels the deterministic fallback separately.

Why the two-agent Strands design matters

Manager Mirror separates the two jobs that make rehearsal useful: one agent stays in the fictional counterpart role and reacts to the manager's wording, while a second agent steps outside the role-play to evaluate observable communication behaviors. The surrounding application handles privacy screening, schema validation, mode disclosure, and deterministic fallback. That separation keeps the interaction more inspectable than a single broad coaching prompt.

Inspiration

Managers are often promoted because they are excellent individual contributors, then expected to handle missed commitments, boundaries, conflict, and unrealistic requests with no safe place to practice. Their first attempt may affect another person’s trust, motivation, and career.

Manager Mirror gives a manager a private rehearsal before the real conversation. It uses fictional scenarios rather than employer records or employee profiles, and it coaches observable communication behavior rather than judging personality.

What it does

A user chooses one of four realistic but fictional moments:

addressing a repeated missed commitment; resetting an after-hours boundary; saying no upward by making the tradeoff explicit; or de-escalating conflict between teammates.

The user writes what they would actually say. The application then produces:

an adaptive fictional counterpart reply; an overall coaching score; six behavior scores: directness, empathy, curiosity, accountability, next step, and composure; strengths and risks grounded in the submitted wording; a stronger version of the opening; and a concise follow-up draft.

The experience is intentionally bounded: practice one conversation, learn what landed, improve the wording, and leave with one concrete next step.

How the agent works

The backend contains a two-agent Strands Agents workflow.

A privacy gate rejects common contact details, URLs, credentials, confidential markers, and other inappropriate input before generation. A counterpart agent role-plays the fictional coworker and reacts to the manager’s actual wording. A separate coach agent scores observable conversational behaviors and returns a structured report. A schema gate validates and bounds the report before it reaches the interface.

The Strands model adapter uses the hosting platform’s AI generation service. A deterministic, tested rubric remains available as a safety and availability floor. The product labels live Strands output and deterministic fallback output separately; deployment availability is never presented as proof that a particular model invocation succeeded.

Why Strands Agents

The problem benefits from distinct agent responsibilities rather than one broad prompt. The counterpart must stay in character and expose ambiguity naturally. The coach must step outside the role-play, evaluate evidence in the user’s language, and produce a stable report contract. Strands provides the orchestration boundary for those two responsibilities while the surrounding application handles privacy, validation, mode disclosure, and failure behavior.

Privacy and safety

Manager Mirror is a practice tool—not an HR system, employee assessment, legal service, or employment decision-maker.

All included situations and people are fictional. Users are told not to enter real workplace or personal information. The application has no contacts database, employee directory, referral flow, messaging integration, or outreach capability. It does not send follow-ups, contact coworkers, invite teammates, request referrals, or take employment action. It provides communication coaching, not legal, medical, HR, or employment advice. A service failure produces a visible error rather than silently showing stale or fabricated coaching.

Technical implementation

React and TypeScript judge-facing application Node/TypeScript backend Strands Agents SDK counterpart and coach agents Hosting-platform AI model adapter Deterministic local evaluation fallback Input privacy screening Structured output and score validation Responsive desktop and mobile interface Automated end-to-end workflow specifications covering the main practice path, scenario specificity, privacy rejection, architecture disclosure, and backend failure

Testing

The fastest judge path is:

Open the live application. Select Missed commitment. Keep the prefilled direct-and-empathetic response. Choose Run the conversation. Inspect the visible runtime mode, counterpart reply, overall score, six behavior scores, strengths, risks, stronger version, follow-up draft, and four-stage trace.

For the privacy guardrail, replace the response with: Email [email protected] about our confidential customer pricing. The application should reject the text before agent execution and show no coaching result.

What I learned

Useful AI coaching is not about generating more advice. It is about creating a controlled rehearsal with clear roles, observable criteria, privacy limits, and truthful degradation. Separating counterpart behavior from coaching behavior produces a more realistic experience and a more inspectable system.

What’s next

The next product step is opt-in voice rehearsal and private user-controlled practice history. Those are future capabilities, not claims about this submission. The current entry focuses on a polished, inspectable text workflow that solves one high-stakes professional problem without collecting employer data or contacting anyone.

Built With

• appdeploy
• aws-strands-agents-sdk
• node.js
• react
• tailwind-css
• typescript
• vite