| name | review-candidates-auto |
| description | Triage the Superposition candidate review backlog unattended ("full auto"). Use when the user asks to auto-review, clear, or work through the awaiting-review queue/backlog, review all candidates for a job without stopping, or batch-process candidates. For decisions the user wants to make themselves one-by-one, use review-candidates-human-loop instead. Requires the company soul to be set up first (bootstrap-company-soul) — assessments are weak without it. |
Review candidates — full auto
Work through the awaiting-review backlog without pausing between candidates. For each one you
gather the FitCheck context with read tools, then let the Master Agent record the decision through
invoke_agent. "Full auto" means you don't stop to ask the human between candidates — it does not
mean skipping Superposition's own approval gates. The Master Agent still enforces them (e.g. it won't
schedule before client approval); when it escalates or needs approval, surface that in your summary.
Prerequisites
- Connected with the right active org (
connect-superposition).
- Company soul is populated (
bootstrap-company-soul). If get_company looks thin, stop and
recommend bootstrapping first — auto-reviewing against a weak profile produces low-quality calls.
Steps
- Scope the org.
get_superposition_context to confirm the active org.
- List active roles.
list_jobs and keep the ACTIVE ones.
- Size the backlog.
get_pipeline_metrics to see how many candidates are awaiting review per
job. Report the total up front and, if it's large, confirm a batch cap (default: 25 per run) before
proceeding so you don't run unbounded.
- Pull the queue per job.
search_candidates({ jobId, stages: ["AWAITING_CLIENT_REVIEW"] }) to
list the awaiting-review candidates.
- Load assessment context.
get_candidates({ candidateIds: [...] }) (batched, up to 20 per call)
for the FitCheck rationale, calibrated traits, and source.category (INBOUND vs OUTBOUND).
- Record each decision via the Master Agent. For each candidate call
invoke_agent, threading
the returned sessionId:
"Review candidate <candidateId> for job <jobId>. Assess fit against the role using their
FitCheck assessment and record the decision with brief reasoning. Follow the standard pipeline
rules; if anything needs client approval or human input, say so instead of forcing it."
- Summarize. Produce one table: candidate (linked), the decision the agent recorded, a one-line
reason, and a flag for anything escalated or left pending. Note how many were processed vs. skipped.
Output
Auto-review — Acme · Senior Backend Engineer (awaiting review: 18, processed: 18)
| Candidate | Decision | Why |
|--------------------------------------------------------|-----------------|--------------------------------------|
| [Diego Prats](https://app.superposition.ai/c/abc-123) | Advanced | Strong distributed-systems match |
| [Mei Tan](https://app.superposition.ai/c/def-456) | Passed | Below the experience floor for L5 |
| [Sam Okoro](https://app.superposition.ai/c/ghi-789) | Escalated | Comp expectation > range — needs client |
Processed 18/18. 1 escalated to a human. None skipped.
Guardrails
- Never fabricate decisions. Report exactly what the Master Agent recorded. If a call fails or is
ambiguous, mark it pending rather than inventing an outcome.
- Respect server gates. If the agent says it needs client approval, can't email an inbound
candidate, or escalates — relay it; don't try to route around it.
- Stay bounded. Honor the batch cap; if more remain, report the remaining count and offer to run
the next batch.
- Use
review-candidates-human-loop when the user wants to make the calls themselves.
