Series 1 of 4 · 35 chapters · 3+1 levels ·

AGENTIC
STACK

Five rungs of autonomy.
Not maximum — optimal.
Pick your autonomy level — then build the stack to hold it.

Start with Topic 01 — Prompt vs. Context → See the full map ↓
35 chapters 3+1 levels
Evidence · Selected

Three receipts. One discipline.

01
10% / 90%
Model vs harness.

The weights are the smallest part of your AI product. The system around them — context, tools, structure — is where it lives or dies.

First principle of the series
02
200K ≠ better
Attention degrades.

Big context windows don't fix bad context. Recall drops long before the window fills. More tokens, worse answers.

Anthropic · long-context studies
03
Tools rewrite reasoning
Before any call.

Tool descriptions shape what the model decides to do. Pick the tools, you pick the agent.

L1·T08 — Tools Shape Behavior
L1 · FOUNDATION10 chapters

Act 01The Stack at Rest

By the end: you can name every layer of context the model sees, and explain why most AI failures live in the layer no one designed.

T01 Prompt ≠ Context A prompt is what you type. Context is everything the model sees. Confusing them is the #1 reason AI products fail silently. READ T02 The Context Window Attention degrades before the window fills. What "big context" doesn't solve. READ T03 The CONTEXT Stack Seven layers: Constitution, Observations, kNowledge, Tracks, Equipment, eXecution, Template. Miss one and the model compensates badly. READ T04 Context Rot More tokens doesn't mean better answers. The signs the rot has started. READ T05 The Constitution Your system prompt is your product definition. Write it like a charter, not a checklist. READ T06 Knowledge at Inference RAG is a context supply chain, not a feature. Retrieval, grounding, and the cost of either wrong. READ T07 Memory Architecture The model doesn't remember. The system remembers for it. What to keep, summarize, and never persist. READ T08 Tools Shape Behavior Tool descriptions change reasoning before any call is made. Pick the tools, pick the agent. READ T09 Structure as Control XML, JSON schemas, templates. Constrain the output space; reliability goes up. READ T10 Your First Pipeline From scattered prompts to an engineered system. The L1 capstone, end to end. READ
L2 · PRACTICE10 chapters

Act 02Context in Motion

The harness in production. Assembly, lifecycle, observability, economics — what runs before every inference call.

T11 What Is a Harness? The 90% of your AI product that isn't the model. Why it decides whether you ship. READ T12 Context Assembly The pipeline that runs before every inference call. Every turn. Under latency pressure. READ T13 The 3C Lifecycle Create. Compress. Cache. The three stages every harness runs, named or not. READ T14 Multi-Turn State What to keep. What to summarize. What to drop. What never to persist. READ T15 The Harness Decision Framework, SDK, or roll your own. Start with the simplest harness that could work. READ T16 Context Observability If you can't see what's in the window, you can't fix it. Minimum viable telemetry. READ T17 Context Economics Every token has a price. Budget. Cache. Survive the compounding math of multi-turn spend. READ T18 Graceful Degradation When context fails, the system shouldn't. Rate caps, timeouts, bad retrieval — designed for. READ T19 Evaluating the Harness Model evals tell you the ceiling. Harness evals tell you what users actually see. READ T20 The Harness-Eval Contract If your evals can't see inside the harness, your product isn't reliable. The loop that holds the bar. READ
L3 · SYSTEMS10 chapters

Act 03The Protocol Layer

Agents as systems. Reliability, coordination, economics. Where the harness becomes a fleet.

T21 The Agentic Shift The model decides what happens next. What changes — and what doesn't. READ T22 The Spectrum of Agency Not maximum autonomy. Optimal autonomy. Five rungs — pick one, don't leap. READ T23 Agent Anatomy Five components. Most agents only need three. What goes wrong when one is missing. READ T24 Architecture Patterns ReAct. Plan-and-Execute. Tree-of-Thoughts. Reflexion. When each pattern earns its complexity. READ T25 The Autonomy Design "Can" doesn't mean "should." Where to give rope, where to tie the knots. READ T26 Multi-Agent Coordination Every handoff loses 37% of context. Flipkart's 5× budget overrun, explained. READ T27 Agent Reliability 90% per step × 10 steps = 35% end to end. The four levers that fix it. READ T28 Agent Governance What the agent can do. Must do. Must never do. Kill switches before launch. READ T29 Agent Economics Single agent: $9. Full harness: $200. The 22× range that decides whether you launch. READ T30 The Living System Models change. Prompts rot. Your architecture must survive both. READ
L4 · FRONTIER5 chapters

Act 04Beyond the Window

What's next. Multi-modal context. Long-horizon agents. The shape of stacks we haven't built yet.

B01 The Protocol Layer MCP + A2A: how agents will talk to the world. Why protocols will outlast frameworks. READ B02 Agent-Native Products Design for agents as users. Interfaces change. Metrics change. Pricing changes. READ B03 Self-Improving Context When agents optimise their own context. The loop that eats the prompt engineer. READ B04 The Trust Architecture Safety at industrial scale. Trust as a system property, not a marketing line. READ B05 The Model-Harness Boundary What shifts when models get smarter. Design for a line that won't sit still. READ

The best agentic teams ship faster because they stopped debugging the model — and started engineering the harness.

Start with Topic 01 → See the full map ↓
Fin · Series 01 of 04 · Thank you for reading
Continue the syllabus. · Three next-hops, in order.
Read next AI PM OS L1·T01 — Why AI PM ≠ SaaS PM Then Evals L1·T01 — Benchmarks ≠ Evals Or jump to Harness L1·T01 — Why Your Agent Fails