Hybrid Execution

Two-agent architecture separating strategic planning (user-facing) from tactical execution (operator-facing). Event-driven state machine with 17+ trigger types — vendor emails, operator updates, stuck-task retries, secure session results. Three-party information asymmetry: the client sees AI, the operator sees AI instructions, AI mediates both sides. Operators rotate in and out; AI maintains continuity. This is what lets us keep completion rates high even when individual automation steps fail — the orchestrator detects failure and reroutes to a human seamlessly.

Matches tasks to operators via skill affinity, SLA pressure, and task-type history. Tracks the blocker chain — is progress stalled on the vendor, the client, or the operator? — and applies different follow-up strategies for each. Check intervals are urgency-adjusted and business-hours-aware. The routing engine is how we turn the 65% agentic rate into 96% — by identifying exactly when and where human intervention is needed and dispatching the right person immediately.

Multi-gate system before any message reaches the user. Entity-level hallucination detection cross-references phone numbers, URLs, and emails against the full conversation history. Predictive sentiment analysis catches conversations trending negative before the user complains. A composite risk score combines execution health signals — task age, operator churn, vendor unresponsiveness — weighted per task type. The quality shield is what prevents automation failures from becoming user-visible failures.

AI follows along with structured playbooks while operators work. Validates progress, catches deviations, provides just-in-time context. Every session generates a stream of events — pages visited, calls made, emails sent — that feed the supervision model. The playbooks themselves evolve: when operators consistently deviate and get better results, that signal feeds back into system improvement. This layer progressively reduces the cost of human involvement while maintaining quality.

Full simulated conversations using LLM-driven blind actors — they don't know the expected outcome. Async events get injected mid-conversation: a vendor email arrives, a phone call goes to voicemail, business hours end. Adversarial personas stress-test edge cases. Thousands of scenarios run against the system before any change ships. This is how we improve the agentic completion rate itself — by identifying failure modes in simulation before they hit production.

Production tasks become simulation scenarios. Failures become harder test cases. Vendor response patterns accumulate into a knowledge graph. Operator performance data improves routing. The loop: production → extraction → simulation → evaluation → improvement → production. Every task the system handles makes the next task slightly more automatable. The system that handles a task today is measurably better than last week's.