Direction 3 · Agent Workflows¶
Agent value isn't "can act autonomously" — it's clear boundaries + least privilege + controllable failure. 3 representative cases + delivery discipline.
Core Capabilities¶
- Planning: ReAct · Plan-and-Execute · Tree-of-Thought · long-context planning
- Tool calls: OpenAI function calling · MCP · JSON schema
- Memory: short-term (session) + long-term (vector) + structured (relational)
- Permissions: read · write · outbound-call three tiers · high-risk human backstop
- Observability: hallucination rate · failure rate · cost · latency · timeout — 5-panel dashboard
3 Representative Cases¶
Case 01 · Bidding Response Agent (Tender Decomposition + Competitive Analysis + Response Draft)¶
- Problem: Chinese IT integrator · tender turnaround 5-7 days · high-value bids missed
- Stack: long-context planning (200-500 page parsing) · scoring extraction · competitive gap analysis · draft generation · human final review
- Boundary: read-only on bidding portal · no auto-submit · final requires human
- KPI: tender turnaround 5 days → 2 days · win rate maintained · response quality review passed
- My role: agent workflow architecture · long-context chunking · scoring extraction rules
Case 02 · Government-Procurement 4-Layer Contract Review Agent¶
- Problem: SOE group with 20 subsidiaries · 600 contracts/month · legal team overloaded · key clauses missed
- Stack: 4-layer review (basic clauses → compliance → commercial → risk) · per-layer LLM prompts · human final review
- Boundary: read-only contracts · outputs suggestions · doesn't modify source
- KPI: legal review workload -60% · mean review time 3 days → 4 hours · missed key clauses -80%
- My role: 4-layer architecture · prompt version management · effect acceptance
Case 03 · CRM 24-Hour Auto Lead-Nurture Agent¶
- Problem: B2B SaaS inbound first-touch 24 hours · competitors stealing high-quality leads
- Stack: lead qualification tiers · personalized templates · meeting scheduler · threshold-based handoff to human
- Boundary: only sends email/SMS · no outbound calls · explicit human handoff at threshold
- KPI: qualified-lead first-touch 24h → 10min · meeting conversion +25% · high-quality lead rate +18%
- My role: agent workflow design · tiering rules · human handoff thresholds
Delivery Discipline (Agent PoC → Production 5 Steps)¶
- Define business closure: what the agent does · what it doesn't · boundary written into the product
- Least privilege: read · write · outbound-call three tiers · high-risk human backstop
- Cost caps: per-call limit + per-user daily budget + global circuit breaker
- Observability ready: hallucination · failure · cost · latency · timeout — 5-panel dashboard
- Canary rollout: 5% → 20% → 50% → 100% · 3-5 days stable per tier
Common Anti-Patterns¶
- ❌ Agent asked to "make decisions autonomously" from day one with no human backstop
- ❌ Only tracking success rate, not hallucination rate — customer complaints within a week
- ❌ Agent allowed to write production data directly — one wrong edit = 100k+ rollback
- ❌ GPT-4 for agent with no cost circuit breaker — single user burns $500/month
Related¶
- Full 20-case portfolio · PDF
- Related directions: Multi-model routing · AI+DevOps