Part IV · Ch. 30 — Agent (Agentic AI)

draft

Hybrid · Multimodal + tools · LLM + reasoning/action loop. A model that plans, acts, observes, and iterates toward a goal — not just responds. Card: ../02-types-of-ai/06-hybrid-neuro-symbolic.kmd.

🎨 Figure F-IV.30.0The perceive-think-act loop. Brief: closed cycle: LLM (think) → tool call (act) → result (observe) → back to the LLM; memory alongside; goal at the top.

Agent — agentic loop

1. Definition and short history

Couples an LLM (ch. 01) to a control loop with tools, memory, and planning. Lineage: ReAct (2022) → AutoGPT → tool use Reflexion to correct course ([metaphor]: not conscious introspection).

  • Multi-agent — several agents cooperating/verifying.

4. Inputs

  • Hardware: LLM inference (GPU); orchestration on CPU.
  • Data: tool-use traces, environments/sandboxes, task eval.
  • Data structures: task graph, vector memory, action history.
  • Systems: tool runtime, sandbox, action queue, observability.

5. Specialized life cycle

Stage Specialization
0 Problem Define target tasks, available tools, autonomy limits
1 Data Traces of successful actions, training/eval environments
2 EDA Scenario coverage, success rate per task type
3 Modeling Base LLM + tool schema + planning/memory policy
4 Training Post-training for tool use; agent RL; fine-tune on traces
5 Evaluation Task completion rate, steps, cost, safety
5.5 Acceptance Red team of dangerous actions, sandbox, limits/permissions
6 Production Loop with real tools; guardrails; human approval for sensitive actions
7 Monitoring Success, cost per task, risky actions, infinite loops
8 Retraining New traces, new tools
9 Governance Autonomy, permissions, action auditing, accountability

6. Capabilities, modes, and modalities

Intellectual + executive: programming, research, software automation, computer operation, multi-step flows. Combines LLM reasoning with action in the digital world; the basis of assistants that do, not just respond.

7. Limits, risks, and ethics

Compounding errors across many steps; irreversible actions; *prompt injection* via tools/content; cost; the need for authorization and guardrails. Autonomy requires auditing and clear limits.

8. State of the art and examples

Claude Code (engineering agent), computer use agents, multi-agent frameworks; trend: more reliable agents, adversarial verification, persistent memory, and cooperation between agents.