Hybrid architectures
Fast reflexes and slow planning in one agent, arranged in layers. The big design question is who gets control when: every layer at once with a referee (horizontal), or up and down a chain (vertical).
Kind: hybrid agent.
Week: 3 (slides 34 to 43, appendix B slides 53 to 55).
Book: Wooldridge 5.2 Hybrid agents (pp. 92 to 94), 5.2.2 InteRRaP (pp. 96 to 98), 5.2.1 TouringMachines (skim).
Techniques on this page: horizontal vs vertical layering, counting the interactions, InteRRaP’s layers and control flow, TouringMachines’ censors and suppressors, comparing the three families.
The problem it solves
Neither extreme is good enough. Deliberative agents plan well but react slowly. Reactive agents react instantly but can’t think ahead or explain themselves.
So combine them: reactive layers handle changes fast, without reasoning; a deliberative layer plans with symbols. Once there are layers, the new problem is meta-level control: which layer is in charge, when? Almost every real agent today, including LLM agents with safety rules around them, is a hybrid.
Words you’ll hear
| Word | Plain meaning |
|---|---|
| layer | one sub-system of the agent, working at one level of abstraction |
| horizontal layering | every layer sees the sensors and proposes actions |
| mediator | the referee that picks between layers’ proposals |
| vertical layering | at most one layer sees the sensors, at most one acts |
| one-pass | control goes up through the layers once; the top one acts |
| two-pass | information goes up, control comes back down |
| knowledge base | a layer’s own store of facts (world, mental or social model) |
| censor / suppressor | TouringMachines rules that block an action / block a percept |
Technique 1: horizontal vs vertical layering
Problem it solves: the two ways to wire the layers together (slides 36 to 37).
- Horizontal: every layer is connected to the sensors and to the action output. Each acts like a small agent and suggests an action; a mediator reconciles them.
- Vertical: at most one layer reads the sensors and at most one produces actions.
- One-pass (filtering): control flows up through each layer and the last one acts.
- Two-pass (abstraction): information flows up, control flows back down, like an organisation where reports go up and orders come down.
Technique 2: counting the interactions
Problem it solves: the book’s argument for why vertical layering is easier to design (pp. 93 to 94), which the slides leave out.
\(n\) layers, each able to suggest \(m\) actions:
- Horizontal: the mediator must handle every combination of suggestions: \(m^n\).
- Vertical: layers only meet at the \(n - 1\) interfaces between neighbours, each with at most \(m \times m\) pairs: \(m^2(n - 1)\).
| \(n\) | \(m\) | Horizontal \(m^n\) | Vertical \(m^2(n-1)\) |
|---|---|---|---|
| 2 | 4 | 16 | 16 |
| 3 | 4 | 64 | 32 |
| 3 | 10 | 1,000 | 200 |
| 4 | 5 | 625 | 75 |
Worked example. \(n = 5\), \(m = 3\): horizontal \(3^5 = 243\); vertical \(3^2 \times 4 = 36\).
With \(n = 2\) they’re equal (\(m^2 = m^2 \times 1\)): there’s only one interface, and every pair of suggestions crosses it. Vertical only pulls ahead from three layers up.
The catch: vertical layering is less fault-tolerant. Control has to pass through every layer, so one failed layer stops the whole agent. Horizontal layering puts a central mediator in every decision, which is a bottleneck.
Technique 3: InteRRaP
Problem it solves: the main worked architecture (Müller 1996; slides 38 to 39). It is vertical, two-pass, with three layers, each paired with its own knowledge base:
| Layer | Knowledge base | Deals with |
|---|---|---|
| cooperative planning layer (CPL) | social model | other agents: joint plans |
| local planning layer (LPL) | mental model | the agent’s own plans and goals |
| behaviour-based layer (BBL) | world model | reactive behaviour |
Below them is the world interface (perception, communication, action). Each layer has two functions: SG (situation recognition and goal activation) and PS (problem solving: planning and scheduling).
- Upward activation: a layer passes control up when the situation is beyond it.
- Downward commitment: a higher layer achieves its goals using the lower layer’s primitives.
A layer may only use its own knowledge base and those below it.
Worked example: a delivery robot on InteRRaP.
| Situation | Path through the layers | Who decides |
|---|---|---|
| obstacle directly ahead | BBL handles it | BBL (a reflex) |
| corridor blocked, needs a new route | BBL → LPL (up); LPL plans; LPL → BBL (down) to move | LPL |
| another robot wants the same lift | BBL → LPL → CPL (up); CPL negotiates; CPL → LPL → BBL (down) | CPL |
Perception arrives at the bottom and actions leave from the bottom. Information goes up until some layer is competent, then control comes back down. That’s “two-pass”.
Technique 4: TouringMachines
Problem it solves: the horizontal contrast (Ferguson 1992; appendix B, book pp. 94 to 96).
Three layers side by side, all fed by perception: a reactive layer (situation-action rules, like “if a kerb is in front and close, change direction”), a planning layer, and a modelling layer (models of other agents, used to predict conflicts). Since every layer can propose actions, a control sub-system of rules arbitrates:
- censor rules stop an action a layer proposed from reaching the effectors;
- suppressor rules stop sensor information from reaching a layer.
InteRRaP vs TouringMachines: both have three layers doing similar jobs. InteRRaP is vertical (two-pass), with a knowledge base per layer; TouringMachines is horizontal, with a rule-based supervisor.
Technique 5: comparing the three families
Problem it solves: the summary table for “compare deliberative, reactive and hybrid” questions (slide 42).
| Deliberative (BDI) | Reactive (subsumption) | Hybrid | |
|---|---|---|---|
| representation | explicit beliefs, goals, plans | none or very little | both, in different layers |
| decision speed | slow | very fast, constant time | fast reflexes, slower planning |
| pro-activeness | strong | weak (percept-driven) | strong in upper layers |
| transparency | high | low | medium; layer interactions are complex |
| weak spot | cost; calculative rationality | no long-term view; no method | complexity; lack of conceptual clarity |
The tension underneath it all is reactivity vs pro-activeness: responding in time vs pursuing goals systematically. Weeks 2 and 3 are three different answers.
Who gets the final word? (slide 40) Should an LLM planner ever override a hard-coded safety reflex? Two defensible answers: the reflex always wins (it’s simple, verifiable and fast; that’s subsumption’s principle), or override only through a transparent channel (logged, human-approved), never on the planner’s own say-so. Either way, the arbitration rule is fixed by the designer, not by the component it constrains.
Today (slide 43): the deliberate-plan-execute cycle became ReAct (think, act, observe); the belief store became agent memory; hand-written plan libraries became LLM-generated plans; InteRRaP’s world model became learned world models.
What to read in the book
| Section | Read? | Why |
|---|---|---|
| 5.2 intro, pp. 92 to 94 | Read | Horizontal vs vertical, and the \(m^n\) vs \(m^2(n-1)\) argument |
| 5.2.1 TouringMachines | Skim | Appendix B of the slides |
| 5.2.2 InteRRaP, pp. 96 to 98 | Read | Slides 38 to 39 |
| 5.2.3 3T, 5.2.4 Stanley | Skip | Not on the slides |
Page numbers are the book’s; in the scanned PDF add 21.
Practice prompt
I'm studying CS5940 Intelligent Agents (Wooldridge, An Introduction to MultiAgent Systems, 2nd ed., 5.2). I want practice on HYBRID AGENT ARCHITECTURES:
- why hybrids exist (reactivity vs pro-activeness) and the meta-level control problem
- horizontal vs vertical layering; one-pass vs two-pass control
- counting designer interactions: m^n (horizontal) vs m^2(n-1) (vertical), and the fault-tolerance trade-off
- InteRRaP: the three layers and their knowledge bases, SG and PS, upward activation and downward commitment; tracing which layer handles a situation
- TouringMachines: its three layers, censor vs suppressor rules
- comparing deliberative, reactive and hybrid architectures on representation, speed, pro-activeness and transparency
- whether an LLM planner should ever override a safety reflex
- mapping the classical ideas to modern agentic AI (ReAct, agent memory, LLM-generated plans)
Please:
1. Use plain, everyday language and explain any term in one short sentence.
2. Give me ONE question at a time, easy first, then harder.
3. Wait for my answer, then mark it and show a model answer with working.
4. If I get it wrong, explain the mistake and give me a similar question.
5. Every 5 questions, summarise my strengths and what to practise.