AI Agent System
From enquiry to response
in minutes, not days
Edvance runs a multi-agent AI system that handles the 70% of school advisory work that is repetitive and rules-heavy — so our consultants can focus on the 30% that requires real judgment.
📩 Enquiry Handler
📨 Outreach Campaigner
📋 Associate Briefer
🔐 Human approval gates
📝 Audit trail
The Foundation
What is an AI agent?
A chatbot completes one turn and waits. An agent keeps running — checking, acting, re-evaluating — until a goal is met or a human intervenes. The loop is what makes it an agent.
The Agent Loop
Observe
Read the current state — user input, tool output, memory, context
Decide
The model chooses the next action — call a tool, ask a question, or finish
Act
Execute that action — call an API, write to a database, draft a document
Repeat
Feed the result back as new context. Go again until the task is done
1. What does the agent know?
Its system prompt, the school's persistent memory, tool results, and conversation history.
2. What tools does it have?
Get school profile, search knowledge base, get history, draft email, update records.
3. Where does the human stay in the loop?
Every irreversible action — sending an email, creating a booking — pauses for approval.
Architecture
Four levels of AI complexity
The Edvance system spans all four levels — from simple structured prompts to a production fleet with memory, tool use, and human gates.
Level 1
Structured Prompt
One call to the model. A well-structured system prompt. No tools, no loop. The foundation everything else builds on.
Example: Ask Edvance portal — answers a school's question from the knowledge base
✓ Built & live
Level 2
Tool-Use Agent
The model is given a set of tools. It decides which to call, your code executes them, the result feeds back in. This is a real agent.
Example: Researcher agent — gets school profile, checks history, searches knowledge base
✓ Built & live
Level 3
Multi-Agent Orchestration
An orchestrator receives a high-level task and delegates to specialist sub-agents with narrow scope. Each agent's output feeds the next.
Example: Enquiry workflow — classifier → researcher → writer, each focused on one job
⚡ New
Level 4
Production Fleet
Persistent memory across sessions. Human-in-the-loop gates for irreversible actions. Append-only audit trail. Every action logged.
Example: School profile persists across all sessions; email approval queue; full audit log
⚡ New
System Design
How the components connect
Every part has one job. The model controls decisions. Your code controls execution. Guardrails are enforced in code — not in a prompt.
Inbound Trigger
Website contact form, manual run by Nathan or Vanessa, or a scheduled cron job. Fires the workflow and returns a job ID immediately.
Orchestrator
Receives the job context, reads the school's persistent memory, breaks the task into sub-tasks, and delegates each to a specialist agent in sequence.
Specialist Agents
Classifier — service type, urgency, associate match. · Researcher — school profile, history, uploaded docs. · Writer — personalised emails or briefing packs. Each has a narrow system prompt and cannot drift into adjacent concerns.
Tool Layer
Real Python functions: get school profile, search knowledge base, get engagement history, update school profile, create engagement record, queue email for send.
Human Gate
LOW — Read operations execute immediately. · MEDIUM — Write operations execute and flag for review. · HIGH — Irreversible actions (send email, create booking) pause and require explicit approval. Nothing goes out without you.
Memory Store
Each school has a persistent Firestore profile. The agent reads this before acting and writes after each run — so it remembers the school across every session.
Audit Trail
Append-only log of every agent action: agent name, tool called, inputs, output, duration, timestamp. Immutable by design — every decision is traceable.
Workflows
Three workflows, built for Edvance
Each workflow maps directly to a real Edvance business process. They run end-to-end — the human only steps in at the gate.
📩
Enquiry Handler
A school contacts Edvance via the website. The agent handles the entire intake process and prepares a personalised response.
1
Classifier agent — reads the enquiry and outputs service type, urgency, complexity, recommended associate, and suggested price range
2
Researcher agent — pulls the school's persistent profile, engagement history, and relevant knowledge base content
3
Writer agent — drafts a personalised response email from Vanessa, using the researcher's findings to make it specific to this school
4
Orchestrator — synthesises findings and adds next steps for Nathan and Vanessa
🔐 HIGH gate — email queued for approval before sending
📨
Outreach Campaigner
Run a personalised outreach campaign to a batch of target schools on a specific topic — without writing each email manually.
1
You provide a list of schools and a topic (e.g. "child protection training")
2
Researcher agent — profiles each school, checks prior contact, finds a personalisation hook specific to that school
3
Writer agent — writes a genuinely personalised email for each school using the research context — not a mail-merge
4
Full batch presented in the approval queue for review
🔐 HIGH gate — entire batch held until you approve each email
📋
Associate Briefer
A workshop is booked. The agent generates a structured briefing pack for the associate who will deliver it — before they arrive.
1
Booking is confirmed with school, associate, service, date, and attendees
2
Researcher agent — pulls school profile, history, and any documents the school uploaded to the portal
3
Briefing writer agent — produces a structured pack: school context, what to emphasise, what to avoid, relevant frameworks, logistics, post-session asks
4
Pack delivered to associate and engagement logged to the school's history
🟡 MEDIUM gate — auto-logged, flagged for Nathan's review
Safety
The human gate — enforced in code
Every tool call is classified by risk before it executes. The guardrail lives in Python, not in a prompt. The agent cannot talk its way past it.
| Tool | Risk Level | What Happens |
|---|---|---|
| get_school_profile | LOW | Executes immediately — read only, no side effects |
| get_engagement_history | LOW | Executes immediately — read only |
| get_associate_profiles | LOW | Executes immediately — static data |
| search_knowledge_base | LOW | Executes immediately — RAG search, no writes |
| update_school_profile | MEDIUM | Executes and logs a flag for Nathan's review |
| create_engagement_record | MEDIUM | Executes and logs a flag for Nathan's review |
| send_email | HIGH | Pauses — added to approval queue, nothing sends until approved |
| create_booking | HIGH | Pauses — requires explicit human approval |
| delete_record | HIGH | Pauses — irreversible, requires explicit human approval |
Philosophy
The 70% Principle
AI should own the 70% that is repetitive, rules-heavy, and judgement-light. The other 30% — the decisions that require relationships, ethics, and accountability — must stay human. Not because AI can't produce an output, but because the consequences of being wrong require a human to own them.
70%
AI owns this
- Classifying school enquiries by type and urgency
- Researching school profiles and engagement history
- Drafting personalised outreach and response emails
- Generating associate briefing packs
- Searching the knowledge base for relevant frameworks
- Logging engagement records and updating school profiles
30%
Humans own this
- Approving or rejecting every email before it sends
- The actual school relationship — the call, the rapport, the trust
- Strategic direction — who to target and why
- Legal and compliance judgements
- Crisis management and sensitive situations
- Final sign-off on any consequential decision
See it in action
The Edvance School Portal includes full access to the agent system — run enquiry handlers, briefing generators, and outreach campaigns from your dashboard.