🤖 Agentic Systems

What Are AI Agents?

An agent is an AI system that takes actions toward a goal over multiple steps, without constant human input. It can use tools, browse the web, write code, and call APIs. This is qualitatively different from chat AI.

• OpenAI's Operator can complete multi-step web tasks autonomously.
• Anthropic's Claude can use computers via 'computer use' API (beta).
• Agents fail differently to chat AI — compounding errors over long horizons.

---

The Agent Architecture

Agents have three components: a planner (LLM deciding what to do), a memory (short-term context + long-term storage), and tools (APIs, code runners, browsers). Understanding this architecture helps you spec what you need.

• LangChain and AutoGen are the two dominant agent frameworks in 2024.
• Tool-calling LLMs (GPT-4o, Claude) select which function to call based on the prompt.
• Agent memory is the hardest part — most fail at tasks requiring >10 steps.

---

Agents in the Enterprise

Enterprise agents automate end-to-end workflows: onboarding new employees, processing invoices, triaging support tickets. The value is in removing human handoffs, not just individual tasks.

• Salesforce Agentforce handles end-to-end CRM update sequences.
• ServiceNow AI agents reduce IT ticket resolution time 60%.
• The ROI multiplier: one agent replaces not one human, but one human + the coordination overhead.

---

Multi-Agent Systems

Complex tasks split across specialist agents: one researches, one writes, one reviews. Multi-agent systems are more capable but harder to debug. They're the future of enterprise AI.

• Google's multi-agent AlphaCode 2 outperforms 85% of competitive programmers.
• Multi-agent debate (agents arguing with each other) improves reasoning accuracy.
• The orchestration layer — which agent does what — is now a core engineering discipline.

---

Agent Safety and Oversight

Agents with tool access can cause real damage: delete files, send emails, make purchases. Every production agent needs scope limits, audit logs, and a human override mechanism.

• 'Prompt injection' attacks trick agents into taking unintended actions via malicious content.
• Best practice: agents operate in sandboxed environments with explicit permission lists.
• All agent actions should be logged with timestamp, input, output, and tool used.

---

When to Use Agents vs Chat

Chat AI: single-turn questions, drafting, brainstorming. Agents: multi-step execution, repetitive workflows, tasks requiring external tool calls. Match the tool to the task.

• If the task requires more than 3 sequential decisions, consider an agent.
• If the task requires real-time data, an agent with search is mandatory.
• Chat AI cost: ~$0.01/query. Agent cost: $0.10–$5 per complete workflow run.

---

Buying vs Building Agents

Off-the-shelf agent products (Zapier AI, Make AI, Devin) handle common workflows. Custom agents are worth building when your workflow is proprietary or high-frequency enough to justify the engineering cost.

• Zapier AI now supports 5,000+ app integrations with natural language triggers.
• Devin (AI software engineer) can implement small features end-to-end from a spec.
• Custom agent ROI: meaningful at >500 workflow runs/month.

---

Measuring Agent Performance

Agent KPIs: task completion rate, error rate, escalation rate (how often human must intervene), cost per workflow. Track these from day one.

• Baseline: what % of tasks complete without human intervention?
• Target: 80% autonomous completion for well-defined, repetitive tasks.
• Cost tracking: agent runs often cost 10–100x more than single LLM calls.