What the agent actually has access to
The registry spans candidate and pipeline actions (moving stages, tagging, flagging, rejecting, prioritizing), screening rule authoring with a live back-test against your own recent interviews, talent sourcing across your pool and the wider candidate network, interview invites, one-off candidate email and open/click tracking, scheduling links and availability, documents and e-signature for offers and contracts, read-only hiring analytics and interview intelligence, the branded email center for reusable templates, job post drafting and publishing, a durable employer memory for standing preferences, recurring scheduled tasks, delegation to sub-agents for multi-step work, saved skills you can name and reuse, and connected accounts across Google Workspace, Microsoft 365, and a hosted remote MCP server for external tools like Claude or ChatGPT.
Each group is its own file in the codebase, reviewed and versioned independently, which is why a new capability shows up in the agent's own self-description the moment it ships rather than waiting on a separate docs update.
A walkthrough: closing out this week's screening
An employer opens the agent Monday morning and says: "how did screening go for the backend role, and get the strong ones booked." The agent calls get_pipeline_overview first, sees 34 applications with 11 still in screening, then calls list_applicants scoped to that job and stage, sorted by overall score.
It reports back six candidates scoring 8 or above, names two of them by strength (one strong on system design, one strong on debugging), and asks whether to advance all six or just the top few. The employer says advance the six. The agent shows an approval card: "Move 6 candidate(s) to Qualified: Priya Nair, Jordan Lee, Aisha Osei (+3 more)?"
Once approved, it offers a scheduling link for the six, generates one scoped to a 45-minute technical round, and stops there, since sending it to each candidate is a separate step the employer has to ask for next.
Boundaries
What the agent will not do without you
Every tool that changes something (a stage move, a rule, an email, a signature request, a job going live) requires your approval before it executes, derived automatically from that tool's operation type rather than left to a per-tool judgment call. Read-only lookups, like reading the pipeline or checking a candidate's history, never need approval and never return a candidate's contact information.
Every candidate id the agent touches is checked against your own employer account first; an id from outside your data aborts the whole action rather than silently skipping it. A rejection email is always offered, never sent on its own, even right after a rejection action completes.
Anything durable it remembers about how you like to hire is stored in a searchable memory you can retire at any time, and any recurring task it schedules is listed and cancelable from the same conversation. Usage is metered against your minute pack, same as interview minutes.
Frequently asked questions
Does the AI recruiting agent make changes to my pipeline on its own?
No. Every tool that writes to your data, whether that is moving a candidate's stage, creating a screening rule, or sending an email, shows you an approval card describing exactly what it is about to do before it does it. Read-only lookups like reading the pipeline or a candidate's history run without approval.
What is the difference between the employer agent and the candidate-facing agent?
They run on structurally separate tool registries. The employer agent acts on your pipeline, jobs, and hiring data with your approval. The candidate agent helps a job seeker search, apply, and track their own applications and never touches your ATS.
Can the agent see candidates' contact information?
Read tools that list or search candidates, such as list_applicants or search_talent_pool, are built to never return email or phone. Contact details only surface through the tools that actually send something, like an invite or an email, and even those require your approval first.
How do I know what the agent is capable of at any given time?
Its own capability description is generated straight from the live tool registry, grouped by category, so it can only describe tools that are actually registered and wired up. There is no separate marketing description that could drift out of sync with what it can really do.
Can I undo something the agent did?
Some actions are reversible through the ATS itself, like moving a candidate back a stage or re-enabling a screening rule. Others are explicitly one-way, like deleting a rule or canceling a scheduled task, and the approval card says so before you confirm.
Related pages
Put the recruiting agent on your pipeline
Talk to the agent about a real screening backlog, a rule you keep meaning to set up, or a batch of offers waiting to go out, and see the approval card before anything changes.