TLDR

The Entire Architecture in 5 Minutes

18
chapters distilled into one page — for the engineer who builds, not browses

1. The Shift

Software finds a new building block every 15 years. Each was driven by pain, not technology.

Procedures 1970s · CALL Objects 1990s · MESSAGE Services 2010s · REQUEST Agents 2025 · DELEGATE ?
Each step solved the previous era's pain. Each introduced new pain. The question mark is the subject of this book.

0.8510 = 0.20

An agent at 85% reliability composed 10 times. Your end-to-end success rate: 20%.

And unlike every previous unit, agents can refuse, improvise, and reason around constraints. The composition problem becomes a governance problem.

The governance gap: Functions have Type Systems, APIs have SLAs, Agents have... Prompts
Functions have type systems. APIs have SLAs. Agents have... prompts. This is the governance gap.

2. The Three-Layer Stack

Every paradigm produces the same three layers. The agent era is no different:

LayerPurposeStandard
ContractTools & capabilitiesMCP (hands)
CommunicationAgent-to-agent messagingA2A (voice)
OrchestrationGovernance & observationHarness (workplace)
The Anatomy of an Agent: Hands (MCP/Tools), Voice (A2A/Communication), Workplace (Harness/Governance)
MCP gave agents hands. A2A gave agents voice. The harness gave agents a workplace.

The agent is the runtime. The skill is the unit. The harness is the operating system. Docker sold for $50M. Kubernetes delivered $34B. Value accrues to the orchestration layer.


3. The AGENT Principles

Five design principles. The SOLID equivalent for agent architecture.

The complete AGENT principles table — five principles, five letters, one word
Five principles. Five letters. One word: AGENT.
PrincipleRule
AAutonomy BoundariesDefine what agents CAN do before what they SHOULD do
GGoverned CompositionEvery connection needs an explicit typed contract
EExplicit ObservabilityTrajectories are source code — record everything
NNarrowest Capable AgentRegex → Function → Prompt → Agent. Use the simplest that works
TTrajectory-Driven ImprovementEvery failure is training data

4. Seven Named Patterns

The Gang of Four named 23 OOP patterns. This book names 7 agent patterns.

PatternWhen to Use
Specialist EnsembleTask spans multiple domains — coordinator + specialists with typed contracts
Verification LoopHigh-stakes output — adversarial verifier, different model, different prompt
Skill MeshCapabilities discovered at runtime from a shared registry
Escalation ChainHospital triage — nurse → doctor → specialist → human
Context FunnelCoordinator curates context slices, not dumps everything (10-50x cost savings)
Consensus ProtocolThree agents vote on high-stakes decisions
Trajectory ReplayGit log + git bisect for agent systems

5. Seven Ways It Breaks

You have hit these. Now you have names for them.

Theory is clean. Production is honest. Intact circle vs shattered circle.
Theory is clean. Production is honest. Seven named failure modes and how to fix each one.
FailureWhat HappensFix
Telephone GameMeaning degrades across handoffsTyped contracts
Echo ChamberSame-model verifiers agree 97%Different model, adversarial prompt
Compound CascadeSmall errors multiply across agentsPer-stage verification loops
Context StampedeDumps everything to every agentContext funnel, curated slices
Guardrail JailbreakAgent reasons around safetyStructural harness enforcement
Cost Spiral$0.50 budgeted → $50 spentGovernor with hard budget caps
State AmnesiaAgent 7 contradicts Agent 2Ledger with full history

6. The Compliance Harness

The central architecture. Four deterministic walls around a probabilistic core. The agent reasons freely inside. The harness enforces compliance outside.

COMPLIANCE HARNESS THE GATE Who can do what? Access control, permission scoping, human approval gates, kill switches. Solves: Guardrail Bypass, Access Control THE LEDGER What happened and why? Immutable audit trail. Every decision recorded. Tamper-evident. 7-year retention. Solves: Audit Trail, State Amnesia, Root Cause THE GOVERNOR How much can it cost? Budget caps, depth limits, circuit breakers. SLA decomposition into per-stage targets. Solves: Cost Spiral, Compound Unreliability THE WITNESS Is the output trustworthy? Independent verification. Different model, adversarial prompt, falsifiable criteria. Solves: Echo Chamber, Compound Cascade AGENT RUNTIME (probabilistic core)
Four deterministic walls around a probabilistic core. Seven regulatory frameworks converge on these same four subsystems.
WallQuestion It AnswersGaps It Closes
GateWho can do what?Guardrail Bypass, Access Control
LedgerWhat happened and why?Audit Trail, State Amnesia, Root Cause
GovernorHow much can it cost?Cost Spiral, Compound Unreliability
WitnessIs the output trustworthy?Echo Chamber, Compound Cascade

SOC-2, EU AI Act, NIST AI RMF, ISO 42001, HIPAA, SR 11-7, and OWASP Agentic Top 10 all independently require these same four subsystems. This architecture is structural, not arbitrary.


7. Where We Are

The 99.9% learn the established framework. The 0.1% create it.
The 99.9% learn the established framework. The 0.1% create it. We are in Phase 2 of 5. The window to define the discipline is now.
Phase 1 Can it work? DONE Phase 2 Can pieces compose? YOU ARE HERE Phase 3 How should they? Phase 4 How at scale? Phase 5 What agents?
Patterns named in Phase 3 become the textbook knowledge of Phase 5. The HTTP spec from 1996 is still in use.

8. The Book in 18 Sentences

One sentence per chapter. The complete argument in 90 seconds.


The staircase has one more step. Nobody has finished building it. Now you have the blueprint.

This page is the map. The book is the territory.

Read the Full Book
← Table of Contents