Pre-launch · join the cohort

Stop studying.
Start shipping.

The Claude Architect exam is for developers who build production agents. Bayesian Knowledge Tracing maps every concept you'll need at the keyboard — MCP design, model routing for cost and latency, hooks, structured outputs, agent loops — to a live probability of mastery. We show you exactly what's weak and drill it until it isn't.

Concepts
175
Mastery threshold
0.90
Domains
5

Domain mastery

Your live mastery map

live
readiness703
concepts mastered94/175
weakest domainContext

Updated after every answer. Threshold to retire a concept: 0.90.

720

The exam for builders

to pass.In 4 weeks.

The exam tests how you ship Claude in production: agentic loops, MCP, hooks, structured outputs, context discipline. So that's exactly what we drill, until your readiness score clears 720. Below it after four weeks? Your next month is on us.

Start the four-week plan

Diagnostic in 20 minutes. No credit card to start.

How we get you there

Diagnostic in 20 minutes. Plan in 1 hour. Pass in 4 weeks.

720 to pass · we don't release you below 800 in mock

Week 1

Diagnostic and plan

  • 20-minute scenario diagnostic across all five domains
  • Initial mastery probability for each of 175 production patterns
  • Daily plan locked onto your weakest 30 patterns first
Week 2

Drill the build patterns

  • MCP tool design, model routing, hooks, structured outputs
  • Concepts retire only at 0.90 probability
  • First scenario-based mock at end of week
Week 3

Transfer to scenarios

  • Archie sessions on every fragile pattern
  • Cross-domain scenarios: agent loops + hooks + JSON in one question
  • Mid-course mock targeting 700+
Week 4

Mock until 800+

  • Full-length scenario mocks every 48 hours
  • We don't release you below 800 in mock
  • Sit the exam with confidence, not hope

Mastery proof

175 concepts. The graph that orders them.

Two artefacts every learner sees from day one: a live concept heatmap across the five domains, and the prerequisite graph that decides what you study next.

Concept space

175 concepts. One probability each.

live
01 · Agentic Architecture27% · 47 concepts
02 · Tool & MCP18% · 32 concepts
03 · Claude Code20% · 35 concepts
04 · Prompt & Output20% · 35 concepts
05 · Context & Reliability15% · 26 concepts
concepts mastered94/175
readiness703
weakest domainContext
mastered ≥ 0.90strongfragilelockedregression

Knowledge graph fragment

Prerequisites are non-negotiable

DAG
Loading diagram...
mastered ≥ 0.90in flightweaklocked

Domain 1: Agentic Architecture, 12 of 36 concepts shown

Mapped to the official exam

Five domains. Thirty task statements. Tracked to 0.90 mastery.

The Claude Architect exam validates that you can make informed tradeoffs when shipping production applications on Claude. The breakdown below comes from Anthropic's exam guide. The engine drills every task statement under every domain, in DAG order, until your mastery probability clears 0.90 on each.

Domain 0127%

Agentic Architecture & Orchestration

Agentic loops, coordinator-subagent orchestration, context passing, hooks for programmatic enforcement, task decomposition, session state and forking. The largest domain on the exam.

7 task statements

Domain 0218%

Tool Design & MCP Integration

Tool-interface design, structured error responses, MCP server integration, tool distribution across agents, the right use of built-in tools. The MCP-shaped half of every production agent.

5 task statements

Domain 0320%

Claude Code Configuration & Workflows

CLAUDE.md hierarchy, custom slash commands and skills, path-specific rules, plan mode vs direct execution, iterative refinement, Claude Code in CI/CD pipelines.

6 task statements

Domain 0420%

Prompt Engineering & Structured Output

Explicit criteria, few-shot prompting, structured outputs via tool use and JSON schemas, validation and retry loops for extraction, batch processing, multi-pass review architectures.

6 task statements

Domain 0515%

Context Management & Reliability

Preserving critical context, escalation and ambiguity resolution, error propagation across multi-agent pipelines, large-codebase strategies, human review workflows, provenance and uncertainty.

6 task statements

The exam itself

The exam, in numbers.

Domains
5
Task statements
30
Scenarios
6
Pass score
720

Scaled 100 to 1000. Scenario-based multiple choice. Source: Anthropic's official exam guide.

How it works

Three layers, one job: get you to 0.90 on every pattern you'll ship.

LAYER 03 · TUTORArchieWhat if a user contradicts the system prompt?LAYER 02 · BKTEngineKP-0420.71KP-1180.94KP-0730.58LAYER 01 · DAGGraphlooptoolresultsubagenthookstop
Layer 01

Knowledge graph

175 concepts. Every one a thing you'll do at the keyboard, not just on the test.

MCP tool schemas, hook ordering, subagent context-passing, model-routing trade-offs, JSON repair patterns — every exam concept is mapped to the production pattern it represents and locked behind its prerequisites. No skipping ahead, no reasoning about hooks before you've nailed the agent loop.

DAG · 175 concepts · 30 task statements · 5 domains

Layer 02

BKT engine

Mastery threshold 0.90. No shipping with a 0.7 on hooks.

Bayesian Knowledge Tracing keeps a live probability that you have actually mastered each pattern. We retire a concept at 0.90 and bring it back the instant a downstream scenario reveals regression. The exam catches a 0.7. So does production.

BKT · per-learner parameters · regression detection

Layer 03

Archie

A Socratic tutor that pressure-tests build decisions.

Archie is built on Claude and constrained to certification content. He never gives the answer. He asks the next question — about your routing decision, your tool schema, your loop termination logic. Every exchange feeds back into the BKT layer, so tomorrow's drills target what you actually struggled to ship.

Claude · graduated hints · misconception detection

Tutor in the loop

Every Archie exchange writes back to the engine.

You won't see the BKT update in real time, but it's happening on every reply. A clean reasoning chain pushes the concept's probability up. A near-miss marks the concept fragile and schedules a return.

Avg. exchanges per concept
2.3
Hint levels available
3
ArchieSocratic tutor · Claude
session · live
hint level 2 of 3·concept KP-042 prompt hierarchy·linked to domain 4

Live in the engine

Concept KP-042 just updated your mastery probability for Prompt Hierarchy from 0.62 to 0.71.

Outcomes

Three architects. Three mastery curves.

Placeholder data · pre-launch cohort

8473.5 weeks

M.K., Solutions Architect

Came in with strong Domain 1 and weak Domain 4. The engine spent 60% of week two on evaluation. Cleared 0.90 on every concept by day 23.

8124 weeks

R.O., Senior AI Engineer

BKT flagged a regression on prompt hierarchy after a Domain 5 mock. Three Archie sessions later, mastery was back at 0.93. Sat the exam two days later.

7915 weeks

A.W., Principal Consultant

Slower start, deliberately. Worked the graph in DAG order. Flat readiness curve until day 18, then the inflection. Final mock 821, exam 791.

From the journal

Notes on the exam, the method, and the model

Frequently asked

The questions architects ask before they buy.

Developers and engineers building production agentic systems on the Claude API. If you're shipping or about to ship code that involves MCP servers, model routing for cost and latency, hooks, structured outputs, or multi-agent orchestration, you're the target. We assume you can already write code; we drill the patterns the exam tests on, in the order that builds correctly.

Five domains across 30 task statements, scored 100 to 1000 with 720 to pass. Questions are scenario-based and presented over 6 case scenarios — the kind of failures and trade-offs you actually hit in production: a tool loop that won't terminate, a hook that should have fired but didn't, a JSON schema that started returning extra keys. Our question bank mirrors that exact structure.

Archie is a Socratic tutor built on Claude. He doesn't quiz you on definitions. He interrogates the decisions you'd make in production: which model to route to, where to put a hook, how to design an MCP tool that won't break under adversarial input, how to recover from a runaway tool loop. Every exchange is grounded in a concept ID and a task statement, and feeds back into the engine that picks tomorrow's drill.

BKT keeps a live probability that you have actually mastered each concept, between 0 and 1. Every answer you give nudges that probability. Once a concept clears 0.90 we stop drilling it and shift attention to your weak areas. It is the same model used by some of the best adaptive systems in education.

Plan for 30 to 45 minutes a day across four weeks. Two of those days will be longer mock-exam blocks. We track your time and adjust the daily plan if you fall behind. Most learners finish in 18 to 22 study days.

No. We are not affiliated with Anthropic and do not administer the exam. You book the official sitting yourself. What we do is get you exam-ready, with a readiness score that maps tightly to actual performance.

One flat fee for full access until you sit the exam. No per-question paywalls, no usage caps on Archie, no separate fees for mock exams. We will publish the exact number on the public homepage once a variant is chosen.

Your study history stays yours. We use it to personalise your plan and aggregate it anonymously to improve the question bank. We do not sell data, we do not train external models on it, and you can export or delete your account at any time.

Start your diagnostic

Stop studying. Start shipping.

20 minutes to your first mastery map. From there, the engine drills you on the patterns you'll actually ship: MCP, model routing, hooks, structured outputs, agent loops.

Begin diagnostic