AI Agent · HR Assistant
AgenticX
One AI assistant the whole company can talk to. It knows the HR policies and does the routine HR work itself.
Durablecrash and deploy safe
Replayableauditable agent runs
1assistant for every employee
4HR jobs in one place
24/7answers and requests
- Users
- Every employee in the company
- Domain
- HR automation
- My role
- Designed and built AgenticX
01
Context
- Employees had HR questions all day, and no single place to ask
- Answers sat in policy documents nobody could find
- Requests like leave and timesheets went through separate HR tools
02
Before and after
Policy questionsSearch documents or ask HRAsk the assistant, answered from the policy
LeaveForms and email follow-upsApplied for in a conversation
TimesheetsManual corrections through HRUpdated by the assistant
Other HR requestsTickets and back-and-forthHandled end to end through one assistant
Failures mid-requestHalf-done work, nobody toldResumes exactly where it stopped
03
The problem
- Policies nobody can findThe answer existed in a document somewhere, so people asked HR instead.
- Routine requests, manual workLeave applications and timesheet changes meant forms, emails and follow-ups for both sides.
- Many systems, one employeeEach HR action lived in a different tool, with its own screens and rules.
- Work that stops halfwayA multi-step request that fails midway (a system timing out, a deploy) used to leave things half done and nobody notified.
04
How it works
1Employee
- Asks in one company chat, in plain language
- Any HR question, any team
2AgenticX agent
- Understands what is being asked
- Decides: answer it, or do it
3Knowledge base
- HR policies and handbooks
- Policy answers come from here
4HR system
- Leave, timesheets and HR queries
- The agent does the work here, on a durable workflow
How a request flows
- 01The employee asks in the company chat.
- 02The agent works out what is needed.
- 03A policy question is answered from the knowledge base.
- 04A request (leave, timesheet, HR query) is handled in the HR system.
- 05The employee gets the answer or a confirmation in the same chat.
05
What we built
- Policy answers
- Leave rules, holidays, benefits, conduct
- Answered from the company's own documents
- No guessing from general knowledge
- Leave and timesheets
- Applies for leave in a conversation
- Updates timesheets directly
- Follows the same approvals as HR
- Any HR request
- Raises and tracks other HR requests
- Separate agents for separate jobs
- New capability = a new agent, not a rewrite
- Never half done
- Every request is a durable workflow
- Resumes exactly where it stopped after a failure
- Every step recorded and replayable
05
Key insight
People trust an assistant that finishes the job. Explaining the leave process is help desk; applying the leave is AgenticX.
06
The hard parts
- Acting safely. Every action follows the same approvals and rules as HR itself.
- Waiting is normal. Requests that wait on approvals or other systems stay safe and replayable.
- The trade-off. Temporal's determinism constraints shape how agent code is written - a learning curve traded for replayable, auditable runs of every request.
07
My part
I designed and built AgenticX: the agent layer, the integrations into company HR systems, and the durable workflow engine that carries each request to completion.
08
What I took away
- Trust comes from finishing the job, not explaining it.
- Plan for failure inside the workflow, not around it.
- Small agents with one job each are easier to add and trust.
Want the detail behind the numbers - architecture, failure modes, what I’d change? Happy to walk through it.
Talk about this project →