Technology Deep Dive — Section 15

Human Agent Integration

AI agents that know when to ask for help. Seamless escalation to humans with SLA enforcement, zero context loss, and automatic resume.

Uniform interface SLA enforcement Suspend-resume

Uniform Interface

In EnGenAI, human agents and AI agents share the same interface. A human is registered with execution_mode='human' and appears in the agent roster alongside AI agents. The system routes tasks based on capability, not species.

Sophi
AI Agent
Execution: autonomous
Autonomy: Level 3
Domain: Backend engineering
Sarah (Legal)
Human Agent
Execution: human
SLA: 24 hours
Domain: Contract review

Live Escalation Simulation

AI Agent (Sophi)
Human Agent (Sarah — Legal)
AI Working
Escalation Trigger
Context Packaged
Human Review
Decision Made
AI Resumes
Task Complete
SLA Timer (24h window)
0%
Monitoring (Task scheduler: 5 min)
Context Preserved (Redis)
Conversation history (12 messages)
Partial code review (3 files)
Compliance check needed: GDPR clause
Agent state: awaiting_legal_review
Zero information loss on suspension
AI → Human Escalation Flow

Escalation Protocol

When an AI agent reaches its autonomy limit or encounters a task that requires human judgement, it escalates automatically. The routing engine matches the task to the right human by role — business analyst, architect, legal, compliance.

1
AI detects escalation trigger
Task exceeds autonomy level, confidence too low, or requires human approval
2
Context packaged
AI saves complete state to Redis breadcrumb — conversation history, partial work, decision point
3
Route to human by role
Matching engine selects the best available human agent based on role, availability, and SLA
4
Human reviews and decides
Approve, reject with feedback, or complete the task manually
5
AI resumes with outcome
On approval, AI continues from exact point of suspension. On rejection, AI adjusts approach.

SLA Enforcement

Every human task has a configurable SLA window. The system sends automatic reminders when time is running out, escalates on breach, and tracks response metrics for accountability.

REMINDER
75%
of SLA window elapsed
BREACH
100%
escalated to manager
MONITORING
5 min
Task scheduler check interval

Suspend-Resume Pattern

When an AI task is escalated, the agent suspends cleanly. All conversation context, working memory, and partial results are preserved in a Redis breadcrumb. When the human responds, the agent resumes from the exact point of suspension — zero information loss.

AI running
Suspended
Human review
AI resumes
Context preserved across suspension boundary — conversation history, partial work, decision rationale

Built on Background Task Scheduling

Human agent task scheduling runs on a background task scheduler with Redis as the message broker. A dedicated periodic scheduler checks SLA windows every 5 minutes, sends reminders, and triggers breach escalations — all without any manual intervention.

GOOGLE A2A PROTOCOL

Every agent (AI or human) exposes an A2A-compatible Agent Card endpoint. This enables standardised agent discovery and inter-agent communication across organisational boundaries.

K8S DEPLOYMENT

Task worker and scheduler run as separate Kubernetes deployments alongside the API and frontend. Each pod is non-root (UID 1001), resource-limited, and monitored via ArgoCD GitOps.

See It In Action

Human agents are first-class citizens in EnGenAI — same API, same routing, same observability. We're onboarding select B2B teams for early access.

Register for Early Access →