Eight agents. Each one has a job, a limit and a log.
Agents run the search. Evidence backs every score. Agents run inside role-specific rules, seat permissions, a frozen scoring rubric and approval gates — and every action they take is recorded.
- Six agents you switch on or pause
- Two always-on system automations
- No candidate released without approval
The roster
Every agent in the system.
Operating state describes how each agent is configured — not a live reading. Activity examples are labelled representative data.
Intake Agent
Always onTurns a submitted role brief into a validated requisition, and moves each new application through parse, hydrate and enrich steps.
Operating state: Always on. Runs on submission; each step is idempotent and safe to retry.
- Inputs
- Role brief, must-haves and constraints
- Job description and CV uploads (PDF only)
- Outputs
- A validated requisition record
- A parsed candidate profile, or a manual-review flag when parsing fails
Controls, approvals and records
- What you control
- Edit any intake field before the role is launched
- Set role intensity: steady, standard or aggressive
- Approval required
- No approval needed to validate and parse. Nothing is released to a client from this step.
- Events recorded
- A processing job row per step, with attempt count and error code
- Audit event for the submission and for every later edit
Representative dataCompleteParsed 3 new applications for Senior Engineer; 1 sent to manual review.
Blueprint Agent
Always onCompiles the validated requisition into a scoreable, versioned rubric with weighted dimensions.
Operating state: Always on. Runs once per role, and again on a re-compile.
- Inputs
- The validated requisition
- Your dimension weighting preferences
- Outputs
- A frozen rubric version
- A written blueprint for review before sourcing starts
Controls, approvals and records
- What you control
- Adjust weights across role fit, evidence, logistics and signal
- Send the blueprint back for changes
- Approval required
- Sourcing does not start until the blueprint is approved. Rubric versions are immutable once scored against.
- Events recorded
- Each rubric version, with who approved it and when
Representative dataWaiting for approvalCompiled rubric v3 for Finance Manager; awaiting blueprint sign-off.
Talent Discovery Agent
You switch it onFinds people who match an open role and adds them to the role's longlist.
Operating state: Switchable per organisation. Off by default; when off, longlists are built by hand.
- Inputs
- The role requirements and location rules
- The talent graph: people already known to the platform
- Talent memory: silver medallists kept from earlier searches
- Outputs
- Longlist entries with a stated reason for each person
- A record in the talent graph that this person was seen for this role
Controls, approvals and records
- What you control
- Switch on, pause or switch off per role
- Change role intensity to change volume
- Approval required
- Cannot contact anyone, and cannot show a person to a client before an admin approves them.
- Events recorded
- Every longlist entry with its stated reason
- A talent-graph record that this person was seen for this role
Representative dataRunningAdded 12 candidates to the Operations Lead longlist this week.
Evidence Agent
You switch it onReads each CV and records evidence for or against each stated requirement, with the source passage attached.
Operating state: Switchable per organisation. When off, applications wait for manual review.
- Inputs
- The CV on file, PDF only
- The role's requirement list and rubric version
- Outputs
- Evidence items with the source passage they came from
- Requirements flagged as unsupported or contradicted
Controls, approvals and records
- What you control
- Switch on, pause or switch off
- Verify or reject each evidence item during review
- Approval required
- Cannot reject a candidate, and cannot release evidence to a client without verification.
- Events recorded
- Who verified or rejected each item, and when
Representative dataNeeds attention2 requirements on candidate A-1042 have no supporting evidence.
Scoring Agent
You switch it onTurns verified evidence into a 0–100 score under one fixed rubric version, deterministically.
Operating state: Always on once evidence is ready. Same inputs and engine version produce the same score.
- Inputs
- Verified evidence set
- The frozen rubric version
- Outputs
- A score run tied to one exact candidate and one exact role
- A per-dimension breakdown with cited evidence
Controls, approvals and records
- What you control
- Hold a score back or return it in the review queue
- Rubric version is fixed for the run
- Approval required
- No score reaches a client until an admin approves the release for that client and role.
- Events recorded
- Each score run with its engine version, rubric version and input hash
Representative dataCompleteScored 8 candidates against rubric v3; 5 above the release threshold.
Pipeline Agent
You switch it onWatches for roles and candidates that have gone quiet and raises them.
Operating state: Switchable per organisation. When off, stalled roles are only spotted manually.
- Inputs
- Stage history and time in stage
- Your service commitments and open decisions
- Outputs
- A flagged role or candidate with what is stalling it
- A task on the right person's queue
Controls, approvals and records
- What you control
- Switch on, pause or switch off
- Set the service commitments it watches
- Approval required
- Cannot move a candidate to another stage, and cannot chase your team without your say-so.
- Events recorded
- Each flag raised, and the task queued against it
Representative dataMonitoringFlagged 1 offer sitting 52 hours without a response.
Coordination Agent
You switch it onRuns the approved contact sequence and offers interview slots from your availability.
Operating state: Switchable per organisation. When off, no messages leave the platform and sequences stay queued.
- Inputs
- Approved message templates and channel rules
- Opt-outs and existing replies
- Your availability windows
- Outputs
- Sent messages logged per channel, with replies in one conversation
- Slot offers and confirmed interviews
- A blocked-with-reason record whenever a rule stops a contact
Controls, approvals and records
- What you control
- Switch on, pause or switch off — pausing stops queued sends
- Approve message bodies and set channel windows
- Set availability windows
- Approval required
- Cannot send an unapproved message body, contact anyone opted out or already in process, or book outside your availability.
- Events recorded
- Every send, reply and block with its reason
- Every confirmed or rescheduled interview
Representative dataScheduled4 slot offers queued for tomorrow; 1 contact blocked (opted out).
Governance Agent
Always onRecords every state change, override, release and access event, and enforces who can see what.
Operating state: Always on. Cannot be switched off.
- Inputs
- Actions taken by agents, staff and client users
- Outputs
- An append-only audit history per role, candidate and decision
Controls, approvals and records
- What you control
- Team roles and seat permissions
- Data visibility settings
- Approval required
- Candidate contact release is a separate permission from candidate visibility.
- Events recorded
- The audit trail itself — appended, never edited
Representative dataMonitoringLogged 46 events across 3 open roles in the last 7 days.
Orchestration
What each agent hands to the next.
Agents do not talk in prose. Each one passes a structured record forward, and the Governance Agent records the handoff.
- 1Intake AgentValidated requisitionBlueprint Agent
- 2Blueprint AgentFrozen rubric versionTalent Discovery Agent
- 3Talent Discovery AgentLonglist entriesEvidence Agent
- 4Evidence AgentVerified evidence itemsScoring Agent
- 5Scoring AgentApproved score runCoordination Agent
- 6Coordination AgentInterviews and repliesPipeline Agent
- 7Pipeline AgentOutcomes and flagsGovernance Agent
Approval gates sit between the blueprint and sourcing, and between scoring and release. Work stops at those gates until a person clears it.
Hard limits
What no agent can do without approval.
These are enforced in the system, not stated as policy.
- Contact anyone — that is the Outreach agent, and it needs its own switch
- Show a person to the client before an admin approves them
- Reject a candidate
- Publish a score to the client — an admin approves every published score
- Send a message body you have not approved
- Contact anyone already in process, opted out, or who has replied
- Contact the same person twice on one channel inside your window
- Cancel a confirmed interview
- Book outside the availability you set
- Change your requirements or your advertised package
- Publish a benchmark that rests on fewer than five closed searches
- Move a candidate to another stage
- Chase your team on your behalf without your say-so
Human approval gate before any candidate reaches your shortlist
Escalation path to a named platform expert inside the workspace
Every override recorded in Governance & Audit
What the agents connect to.
- Blueprint CompilerCompiles that intake into the versioned rubric agents score against.
- Evidence GraphLinks every requirement to the CV line or note that supports it.
- Scoring EngineProduces 0–100 scores from cited evidence under a fixed rubric version.
- Decision WorkspaceWhere you compare, approve, reject and advance candidates.
- Governance & AuditRecords every state change, override and access event.
Ready to see it
Open a role and switch the agents on.
You choose which agents run, what they may send, and what must be approved before a candidate reaches you.