Ana Stanojević

Reference behavior map for the agent system; decisions ship incrementally behind one policy surface.

System behaviors

• strategy service — async crawlers/feeds, normalize/dedupe, rank, enqueue.
• queue manager — opportunity queue, lease/dequeue, priority, cross-job budget hints.
• runtime controller — per-job state machine, iteration budget, routes agents, terminal decisions.
• memory store + retrieval layer — structured facts, snippets, drafts, traces, outcomes, company records; similarity / filters for planner and writer.
• policy engine — submission gates, confidence thresholds, escalation rules, hold-for-approval.
• browser executor — instrumented navigation, DOM extraction, apply-path classification.
• application executor — draft/export/email packet, optional form fill, submit only via policy.
• outcome tracker — ingest replies/silence; emit events for policy and strategy.

Core state objects

• Opportunity — source id, normalized posting, rank score, queue position, dedupe key.
• ExecutionContext — job id, budget remaining, policy snapshot, apply-path enum, trace id.
• EvidenceBundle — CV facts, retrieved snippets, prior successful patterns per requirement cluster.
• DraftRevision — versioned draft, critic verdicts, grounding map.
• DecisionRecord — terminal action, confidence, policy version, human-approval flag.
• OutcomeEvent — channel, label (reply/reject/silence/interview), timestamps for adaptation.
• PolicyConfig — thresholds, allowed auto-submit surfaces, company blocklist, rate limits.

Execution policies & safety

• No submit without passing critic grounding; high-risk paths require human approval.
• Rate limits and per-source caps; explicit blocklist for companies/domains.
• Budget caps per job and per rolling window; starvation limits for low-priority queue items.
• Browser session isolation, credential vaulting, audit log of actions.
• Outcome feedback updates ranking weights and retrieval priors—versioned policy rollouts.

Migration from simpler build

1. Split single runtime into queue + controller; persist Opportunity and ExecutionContext.
2. Add retrieval API behind planner; backfill memory schema (facts, drafts, outcomes).
3. Introduce policy engine as single gate for submit/hold/escalate; route all terminals through it.
4. Stub browser executor behind feature flag; ship DOM extract + apply-path detect before full automation.
5. Wire outcome tracker to strategy ranker for closed-loop adaptation.