Agents en production — Automatisation, supervision et cas d'usage réels

De demo à production

En démo : Une requête = agent répond. En production : Centaines par jour, 24/7, avec risques réels.

3 patterns de production

Pattern 1 : Supervised Automation

Agent runs, human reviews before action.

Workflow:
1. User submits request
2. Agent runs (max 5 iterations)
3. Agent output → Approval queue
4. Human reviews (2 min max)
5. If approved → Execute final action
6. If rejected → Agent refines & resubmit

Best for : Financial decisions, customer communication, compliance.

Pattern 2 : Autonomous with Monitoring

Agent runs and acts, monitoring for anomalies.

Workflow:
1. Agent runs autonomously
2. Execute result immediately
3. Monitor : Did it work ? Errors ?
4. If error → Escalate to human
5. If success → Logged

Best for : Repetitive tasks (support triage, data ingestion), high-volume.

Pattern 3 : Hybrid (Most common)

Low-risk tasks = autonomous. High-risk = supervised.

If (task_risk_score < 3 and not_involves_pii and not_critical) {
    autonomous()
} else {
    supervised_with_human_approval()
}

Cas d'usage réels

Finance : Invoice processing agent

Tools:
- ocr_invoice() → extract {vendor, amount, date, items}
- validate_against_po() → {valid, discrepancies}
- create_payment() → {payment_id, status}

Workflow:
1. Invoice arrives
2. Agent OCRs it
3. Validates against PO
4. If match → Auto-approve (if <$10k)
5. If discrepancy → Escalate to AP team
6. Logs every step

Support : Ticket routing agent

Tools:
- analyze_sentiment()
- classify_issue()
- get_kb_suggestions()
- create_ticket()
- assign_to_team()

Workflow:
1. Customer message arrives
2. Agent analyzes tone + issue
3. If FAQ-able → Auto-reply with KB
4. If needs human → Create ticket + route to team
5. Measure : % resolved auto vs routed

HR : Resume screening agent

Tools:
- extract_resume()
- match_against_jd()
- score_candidate()
- send_status_email()

Workflow:
1. Resume submitted
2. Agent extracts key info
3. Scores against JD (match %, red flags)
4. If score > 75% → Pass to recruiter
5. If score < 40% → Auto-reject (email candidate)
6. Middle range → Recruiter reviews

Production checklist

Infrastructure

[ ] API for agent to call tools (rate-limited, authenticated)
[ ] Database logging every action
[ ] Error handling + retry logic
[ ] Monitoring + alerting (if agent fails)
[ ] Fallback : if agent crashes, escalate to human

Governance

[ ] Approval workflow for high-risk actions
[ ] Audit trail (who/what/when/why)
[ ] SLA for human review (if supervised)
[ ] Kill switch : ability to pause agent instantly

Monitoring

[ ] Dashboard : % automated vs supervised
[ ] Error rate tracking
[ ] User satisfaction (did agent help ?)
[ ] Cost tracking (token usage)
[ ] Iteration count (is agent looping too much ?)

Performance metrics

Metric	Target	Red flag
First-time resolution	>70%	<50%
Avg tool calls/request	3-5	>10 (looping)
Escalation rate	20-30%	>50%
Human review time	<2 min	>5 min
Error rate	<1%	>3%

Common failure modes

❌ Failure 1 : Agent gets stuck in loop

Calling same tool repeatedly without progress.

Fix : Track tool call history. If repeated 3x → escalate.

❌ Failure 2 : Tool error not handled

Tool returns error → Agent ignores → produces hallucination.

Fix : Wrap all tool calls in try-catch. Tell Claude "tool failed."

❌ Failure 3 : Context explosion

Too many messages → token limit → Claude cuts off.

Fix : Summarize old messages. Keep only recent context.

❌ Failure 4 : Unintended side effects

Agent's action broke something (sent wrong email, etc.).

Fix : Always supervised for actions with side effects.