The Full System — Positions, Thinking, Memory & Patterns
Every concept on this page in one diagram. The five autonomy levels feed into the agent's ReAct thinking loop — powered by three layers of memory — all wrapped in human oversight. Multiple agents can be arranged using the four patterns below.
Hover over nodes and patterns to explore. Scroll down for interactive deep-dives into each section.
What Does Each Click on the Dial Actually Mean?
Think of each level as a different travel arrangement — from a website that shows you flights to an agent who manages every trip you'll ever take.
What helped me most was sticking to one example the whole way through. One trip, one booking, one agent — from the moment it receives the goal to the moment the flight is confirmed. That same progression maps exactly to real agentic builds. The rawweights.com rebuild used every position on this dial: scripted page audits, AI-checked suggestions, multi-agent design, fully autonomous deployment of low-risk pages. The dial isn't theoretical — every notch describes a real decision you make when you ship.
Position 1
Suggest
Scripted / Rule-BasedYou enter your travel dates. The search engine lists twelve flights matching your filters. It does nothing else. You compare, you decide, you book it yourself.
Travel equivalent
A flight comparison site. It reads the schedule, you make the call.
Position 2
Approve
Bare LLM / ChatbotA travel assistant finds the best available flight for your trip, cross-checks your loyalty account, and sends you a message: "Confirm this booking?" Nothing is charged until you reply yes.
Travel equivalent
A concierge who does all the research and hands you the pen to sign.
Position 3
Delegate
Agentic WorkflowYour travel agent has standing instructions: book economy, always grab a window seat, hotel within walking distance of the venue. She handles every standard trip. She calls you only when something unusual comes up — a cancellation, a price spike, an upgrade opportunity.
Travel equivalent
A corporate travel desk that works within your company policy and escalates edge cases.
Position 4
Monitor
Single AgentYour entire travel calendar runs independently. Hotels, transfers, seat upgrades, check-in reminders — all handled. You receive a weekly digest. You can step in anytime, but you rarely need to.
Travel equivalent
A full-service travel manager for a mid-sized company — responsible, reporting in, acting alone.
Position 5
Automate
Multi-Agent SystemAll recurring trips run without you. A scheduling agent reads your calendar, a booking agent handles reservations, a disruption agent reroutes you when a flight is delayed. You only hear from the system when something it cannot resolve on its own requires a decision.
Travel equivalent
A 24-hour travel operations centre with specialized teams working in parallel on your account.
Higher autonomy isn't automatically better
Fully automating your routine Monday commute is very different from fully automating your company's international conference travel. The right position depends on the stakes involved, how easily a mistake can be undone, and how well you understand what the agent is doing on your behalf.
Autonomy Without Control Is Not Freedom — It's Exposure
Autonomy is only one axis. The other is how tightly you can see, adjust, and limit what's happening. Where those two axes intersect is where your actual risk lives.
The Agent with the Company Card
Your travel agent has full booking authority, no spending ceiling, and sends you updates only when she feels like it. She is booking everything — you won't see the bill until after the trip.
Verdict: Dangerous
The Trusted Agent with a Budget Ceiling
She handles your full travel calendar — but has a £500-per-booking limit, sends you a nightly summary, and you can cancel anything confirmed in the last hour.
Verdict: Enterprise sweet spot
The Flight Search Website
It shows you results. You click around. It doesn't act, and you're not really watching it — because there's nothing to watch. Limited risk, limited value.
Verdict: Safe but limited
The Check-In Desk Agent
Every action requires your passport, your signature, a supervisor code. A seat change generates a paper trail. Nothing happens without you present. Appropriate where errors carry real consequences.
Verdict: Regulated high-stakes environments
← Low autonomy
Autonomy axis
High autonomy →
Goal → Reason → Act → Observe → Repeat
An agent isn't following a checklist. It reasons in a loop — constantly checking whether its last action moved it closer to the goal, and whether the situation changed in a way that should alter what comes next.
Scenario — booking a 3-night trip to Lisbon
Step 1 — Receive the Goal
"Book a 3-night trip to Lisbon for two people, leaving next Thursday. Budget under £800 total, hotel near the city centre."
The agent now has a clear task, constraints to work within, and a starting point for its planning.
🔄
It loops
Each action triggers an observation that informs the next decision. Not a checklist — a feedback cycle.
🔧
It uses tools
The agent searches availability, checks prices, reads your loyalty account, and books — not just generates text about booking.
🎯
It recovers
When the hotel is full, the flight is cancelled, or the price spikes — it doesn't stop, it replans.
Three Layers of Memory, Three Lengths of Time
An agent's ability to stay useful over time depends entirely on which kind of memory it can access — and how long that knowledge survives after a conversation ends.
Working Memory
The sticky note on the desk right now
Right now, the agent knows: 3 nights, Lisbon, two people, Thursday departure, £800 budget. That's it. This is active, immediate knowledge — the facts it is currently working with for this single request.
Lifespan
This request only
What it holds
Destination, dates, budget, number of travellers
When it disappears
The moment this task completes
Session Memory
The notepad from today's conversation
Everything said during this planning session. You mentioned wanting an aisle seat. You said to avoid the 6am departure. You asked for a hotel with a gym. The agent remembers all of it — until you close the chat and start a new one tomorrow.
Lifespan
All turns within one conversation
What it holds
Seat preference, departure time, hotel amenities, mid-session changes
When it disappears
When you close the session
Long-Term Memory — Two Kinds
Your travel profile and your travel diary
📇 Semantic — The Profile
Standing facts with no expiry date. "Always books an aisle seat." "Vegetarian meal option on long-haul." "Prefers the same hotel chain for loyalty points." These apply to every future booking, forever, until you change them.
📔 Episodic — The Diary
Time-stamped events. "In April 2024, a last-minute rebooking cost £340 — the traveller asked us to keep a buffer." "In November 2024, the traveller mentioned they now use a mobility aid." These happened at a specific moment and change how the agent handles future requests.
Together, these make the agent feel like it already knows you — because knowledge from every past trip was saved and is retrieved whenever it becomes relevant.
Are You the Co-Pilot or the Control Tower?
There are two fundamentally different ways to stay involved with a running agent — and one of them doesn't mean you've stopped paying attention.
Human-in-the-Loop (HITL)
The Co-Pilot
You are inside the decision chain. Before the agent confirms any booking, it stops and sends you the details to approve. Nothing is charged, no confirmation is sent, no itinerary is locked — until you say yes.
Best for situations like:
- → First-time use of a new agent you haven't calibrated yet
- → Any booking above your normal spend threshold
- → Non-refundable tickets or hotel deposits
- → Trips with complex multi-leg itineraries
Human-on-the-Loop (HOTL)
The Control Tower
The bookings are happening. You are watching the dashboard. You can intervene — and you will if something looks wrong — but the agent doesn't pause to ask permission for every move. Your job is oversight, not approval.
Best for situations like:
- → Routine recurring trips you've seen many times before
- → High-volume bookings where per-approval would create a bottleneck
- → An agent that's been running reliably for months
- → Time-sensitive bookings where delay has a cost
Four situations where you should always require a check-in — regardless of which mode you're in:
💰
Before money moves
Any payment above your set threshold — no exceptions
📣
Before external comms
Booking confirmations sent to clients or third parties
❓
When it flags uncertainty
Two options look equally valid and the agent can't decide
🚫
Non-refundable actions
Non-flexible fares, hotel deposits, airline tickets already ticketed
Four Ways to Organise Who Does What
One agent can handle a simple trip. But complex travel — 12 people, 6 cities, shifting schedules — is often better split. These four patterns describe how you divide responsibility.
Pattern 1
The Router
Triage — no action taken
The hotel concierge desk. When a guest asks something, they don't handle it themselves — they figure out who can, and send you there. The router's only job is to read the request and direct it to the right specialist.
A travel query arrives → Router reads it → Sends it to the "flights" agent, the "hotels" agent, or the "disruptions" agent based on what's being asked.
Pattern 2
The Orchestrator
Coordinates without doing
The tour operator who plans a 10-day itinerary. She doesn't drive the coach, manage the hotel, or run the tours herself — she coordinates the specialists who do, and assembles the pieces into one coherent trip.
Book a conference trip → Orchestrator assigns: flight agent handles air, hotel agent handles rooms, transfer agent handles ground. Results roll up and get assembled into one itinerary.
Pattern 3
The Supervisor
Monitors and intervenes
The on-trip coordinator who actively watches the whole journey in real time. When your connecting flight is delayed by 90 minutes, she reroutes the airport transfer and calls ahead to the hotel — without anyone asking her to.
Flight lands late → Supervisor detects the delay → Immediately rebooks the transfer, updates the hotel arrival time, and alerts the traveller — all before they've left the gate.
Pattern 4
The Reflection Loop
Produce → critique → improve
A travel planner and a quality checker working together in a loop. The planner drafts the itinerary. The checker reviews it: "Day 3 has a 4-hour gap but the transfer alone takes 90 minutes." The planner revises until the checker has no more objections.
Draft itinerary → Critic agent finds a timing clash → Planner revises → Critic approves → Itinerary is finalised and sent to the traveller.
Five Ways an Agent Fails
Every failure mode here has a real-world travel equivalent — and they all appear in the wild. Recognising the pattern early is the difference between a minor redirect and a missed flight.
Failure 1 — Stale Knowledge
Booking a hotel that closed six months ago
The agent searches for accommodation, finds a well-rated option near your venue, and books it. It doesn't know the hotel went through a major renovation and closed its doors to guests six months ago. It followed its information perfectly — the information was just wrong.
This is the equivalent of GPS routing you down a road that was closed in the last map update. The system is technically working. The world has moved on without it.
Prevention: Agents need access to live, current information — not just what they were trained on. Build in expiry dates for knowledge, and ensure critical checks (like hotel availability) happen in real time at the moment of booking.
Failure 2 — Overconfident Assumption
Booking the wrong London airport
You ask the agent to book a flight from London. You mean London City — it's 15 minutes from your office. The agent assumes Heathrow, which is how most London flights are listed. Tickets issued. Your departure is 40 miles in the wrong direction.
The agent didn't flag the ambiguity. It filled in the gap with its most common pattern and proceeded as if it were certain. A well-calibrated agent would have asked.
Prevention: Build calibrated uncertainty into the system. When a key detail is ambiguous, the agent should ask — not guess. Confirm before you commit to anything non-refundable.
Failure 3 — Infinite Search Loop
Still searching for the perfect flight while yours sells out
The agent keeps searching for a better combination of price, timing, and stops. Option A is slightly too expensive. Option B has a short layover. Option C is good but maybe something better will appear. After 90 minutes of recalculating, your preferred flight sold out at the price you would have paid immediately.
This is a reasoning loop with no exit condition — the agent never hit "good enough" and kept going indefinitely instead.
Prevention: Set a maximum number of reasoning steps. If the agent hasn't converged after a defined number of attempts, it should return its best option found so far — not keep searching forever.
Failure 4 — Precise Execution of Wrong Instructions
Perfect itinerary, wrong destination
You asked for a "hotel near the conference." The agent books one near the conference registration address — which is the venue's administrative office, in a suburb 45 minutes from where the actual sessions take place. The agent did exactly what you said. You said the wrong thing.
An agent optimises for the instruction it received, not the outcome you intended. Vague inputs produce precise-but-wrong outputs.
Prevention: Specific constraints with clear definitions of success. Before the agent starts, define what "near the conference" actually means — a postcode, a radius, a specific area of the city.
Failure 5 — Wrong Tool for the Job
Using the flight search to book a dinner reservation
The agent is given access to a flight and hotel booking tool. You ask it to also reserve a table at a restaurant for your client dinner. It tries to process the restaurant request through the hotel booking system, fails, loops, and reports the restaurant as "unavailable" — because it was querying the wrong service entirely.
An agent without the right tools will improvise with what it has. The result looks like failure, but the underlying problem is a tool mismatch.
Prevention: Match tools to tasks deliberately. Give the agent exactly what it needs for its specific scope — no more, no less. If you expand the task, expand the toolset to match.
Why chained steps multiply risk
If your agent is 95% accurate at each step and runs through six reasoning steps to plan your trip — checking flights, availability, prices, transfers, hotel, and confirmation — you don't end up with 95% accuracy overall. You end up with roughly 74%. A small error at step two compounds through every step that follows it. Shorter chains with human checkpoints outperform long autonomous pipelines for anything high-stakes.
Don't Hire a Travel Agency to Book One Taxi
Before building any agent, ask whether every piece of it is actually necessary — or whether you're adding complexity out of habit.
Fewest agents that can make the booking
One agent handling flight search, hotel, and confirmation beats three specialist agents in parallel before you understand how they fail.
Intelligence matched to the trip's complexity
Booking a regular weekly flight to the same city doesn't need a reasoning model. A multi-stop trip with policy constraints does. Right-size the model to what's actually being decided.
Every booking tool earns its place
Don't give the agent an expense management integration until it's actually failing on expense reporting. Each new capability adds a new surface for something to go wrong.
Access limited to what the trip actually needs
Flight search and email confirmation — those two tools. Not your full calendar, not your expense system, not your entire contact list. Each extra permission is a surface for a failure you didn't plan for.
Four questions before you build your travel agent
The same questions apply to any agentic system — the travel framing just makes them easier to reason about
Could a simple booking site handle this reliably?
If the trip is always the same route, same hotel, same day of the week — you don't need an agent. A standing order or a direct booking tool is faster, cheaper, and less fragile.
Does this trip actually need multiple specialist agents, or can one handle it?
A multi-agent system adds handoff points and failure modes between agents. A single well-scoped agent handling the whole booking is often faster and much easier to debug when something goes wrong.
Which tools does it actually need access to?
If the agent only needs to search flights and send a confirmation email — give it those two tools. Not your full calendar, not your expense system, not your entire contact list. Each additional tool is a surface for error.
What is the simplest setup that could work?
Start there. Book the flight with the simplest possible agent. Run it on one trip. Watch what breaks. Add complexity only when the simple version demonstrably can't do what you need.
🚕
Start with the taxi
If you need to get from the airport to your hotel, you book a taxi — not a full travel operations centre with a fleet manager, a route optimiser, and a customer success team. Agents built to handle tasks that didn't need them add overhead without adding value, and they introduce new ways to fail. Prove the simple version works first. Then scale the one piece that actually needs to grow.
Public Sources & Further Reading
ReAct: Synergizing Reasoning and Acting in Language Models
Yao et al., 2022 — Google Research & Princeton · arXiv:2210.03629
Building Effective Agents
Anthropic Research — Patterns for autonomy levels, human oversight, and agent design
Agents
Chip Huyen — AI agent architectures, the ReAct loop, and autonomy spectrum
Human-in-the-Loop & Human-on-the-Loop
AWS / Azure ML documentation — Standard industry terminology for human oversight modes