People Operations Agents
Task-specific chat prototypes and a structured onboarding-plan builder for People teams.
- Next.js
- AI
- People Operations
Under the hood
Technical architecture
These prepare drafts. They do not execute actions in an HRIS, ATS or workflow platform. Chat history and draft plans are held in browser component state.
Task interface
Next.js · React dialogs
Selects the agent, collects context, preserves failed submissions and presents replies or onboarding phases.
Server routes
/api/agents/chat · /api/agents/onboarding
Validates inputs, applies agent instructions and requests a response. The onboarding route parses the returned JSON into a plan.
Provider boundary
Server-side AI adapter
Tries configured Deepgram, OpenAI-compatible Kimi, OpenAI and Gemini paths through bounded requests. Credentials stay server-side; availability depends on configuration and the provider.
Output
Conversation · structured plan · exports
Returns drafts for human review. The onboarding UI can export Markdown or CSV.
Failure and fallback path
When services fail, the UI offers retry and a clearly labelled prewritten onboarding example. That example is static content and does not demonstrate live AI availability.
One request, end to end
- Choose a task and submit context with bounded conversation history.
- The route composes instructions and calls the configured response adapter.
- The interface displays a draft, or retains inputs and offers recovery.
Why it is built this way
- Keep provider credentials out of the client.
- Preserve failed inputs and avoid duplicate conversation turns on retry.
- Keep draft preparation separate from permission to execute a real People process.