Français

🧩 Multi-agent architecture

Applied AI · Advanced level · Session 4 · Yann Isola
Preparation Claude Certified Architect

1 Task breakdown viewer

Choose a complex task (or enter your own), then compare how each strategy breaks it down into agents. Observe the structure, latency, and risks unique to each approach.

The three decision criteria

CriteriaQuestion to askIndicative threshold ⚠
Complexity More than 5–7 heterogeneous steps? Does one prompt cover everything?> 5–7 steps → decompose
Token budget Does the necessary context saturate the window (~60–70%)?Predictable saturation → break down
Specialization Would specialized prompts measurably improve quality?Quality gain > orchestration overhead
💡 Architect's reflex: the most robust architecture is the one that does not exist. If the task fits in a context with acceptable quality — don't break it down. The volume of executions is handled by parallelization of instances , not by decomposition.

2 Escalation Flow Builder

Design an escalation policy: add rules (trigger → pattern → priority), then test it against scenarios. HITL = Human-in-the-Loop , human in the loop.

➕ New escalation rule

📋 Current policy

No rules. Add or upload the sample policy.

🧪 Test a scenario

Certification pitfall: a confident hallucination has self-reported confidence high . The confidence threshold alone does not catch it — hence the need for the independent “outside perimeter” trigger (defense in depth). Test scenario 3 with a policy containing only one trust rule to see this.

3 Circuit breaker simulator (software circuit breaker)

An orchestrator calls a subagent every 1.5 s. Toggle the failure switch and observe the state transitions: CLOSED → OPEN → SEMI-OPEN . Threshold: 5 consecutive failures · Cooling: 8 s (demonstration values ​​⚠).

FARM
normal calls
OPEN
immediate rejection
SEMI-OPEN
a test call
FARM
Consecutive failures: 0 / 5
Calls: 0 · Succeeded: 0 · Failed: 0 · Rejected (open circuit): 0
💡 Educational scenario: start healthy, switch the failure, count the 5 failures until opening, observe the instantaneous rejections (fail fast ), wait for cooling, then switch back to healthy during the semi-open state to see the closure — or leave the fault to see the return to OPEN.

Log (with correlation ID)