TUTOR

M17 · Level III — AI Systems Engineering

Agentic Systems & Tool Use

Engineer agentic systems in which a model selects or proposes actions, invokes tools, consumes observations, maintains state and iterates toward a declared objective.

Topics

  • Goal
  • State
  • Plan
  • Tool
  • Observation
  • Verification
  • Memory
  • Recovery
  • Decision
  • Action

Technical concepts

  • 01The Agentic System as a State-Transition Process
  • 02Goals, Constraints and Success Conditions
  • 03Tool Interfaces as Typed Contracts
  • 04Tool Taxonomy
  • 05Tool Selection as a Routing Problem
  • 06Planning and Task Decomposition
  • 07Reactive vs Deliberative Agent Loops
  • 08State Representation
  • 09Memory: Working, Episodic and Persistent State
  • 10Memory Write Policies
  • 11Verification as a First-Class Agent Component
  • 12Preconditions, Postconditions and Invariants
  • 13Retries, Idempotency and Duplicate Actions
  • 14Failure Recovery and Replanning
  • 15Uncertainty in Agent Decisions
  • 16Function Calling and Structured Action Generation
  • 17Code and Computation Tools
  • 18Retrieval + Tools: Knowledge vs Action
  • 19Classical Planning Foundations
  • 20Language Models as Planners: Strengths and Limits
  • 21Search Over Tool Actions
  • 22Multi-Agent Systems
  • 23Agent Communication Protocols
  • 24Human-in-the-Loop and Approval Boundaries
  • 25Prompt Injection and Tool-Use Attacks
  • 26Policy Engine and Capability Boundaries
  • 27Sandboxing and Execution Envelopes
  • 28Agent Trajectory Logging
  • 29Agent Evaluation Metrics
  • 30Agent Benchmark Design
  • 31Termination, Loops and Budget Control
  • 32Observation Validation
  • 33Memory Retrieval and Context Selection
  • 34Planning, Deliberation and Hidden Computation
  • 35Critic, Verifier and Executor Roles
  • 36Agent Memory + Retrieval + Tools
  • 37Simulation as the Safe Training Ground
  • 38Quantum Technology Bridge: Agents and Hybrid Decision Loops

Practical implementation

  • 01Laboratory I: Typed Tool Simulator
  • 02Laboratory II: Single-Agent Closed Loop
  • 03Laboratory III: Planning and Replanning
  • 04Laboratory IV: Agent Memory System
  • 05Laboratory V: Verification and Recovery Engine
  • 06Laboratory VI: Retrieval-Augmented Research Agent
  • 07Laboratory VII: Planner-Critic-Executor Architecture
  • 08Laboratory VIII: Agent Security Stress Test
  • 09Laboratory IX: Agent Evaluation Harness

Tools and technologies

  • Linux VM/container
  • deterministic tool simulators
  • structured action schemas
  • traceable state
  • evaluation harness
  • failure injection

Project

End-to-end agent with typed tools, policy, memory, verification and evaluation harness

Expected learning outcomes

  • 01Can formally specify goal, state, action, observation and termination.
  • 02Can design typed and validated tool interfaces.
  • 03Can enforce capability and permission boundaries.
  • 04Can build a single-agent closed loop with replayable trajectories.
  • 05Can implement planning, replanning and recovery.
  • 06Can design working and persistent memory with provenance.
  • 07Can verify postconditions and handle silent failures.
  • 08Can evaluate agent trajectories using success, cost, recovery and security metrics.
  • 09Can reproduce an agentic research result at manageable scale.
  • 10Can defend the system's limits and safety boundaries in a technical viva.

Assessment

  • 01Daily practice
  • 02Laboratories
  • 03Research reproduction
  • 04Integrated project
  • 05Technical viva
  • 06Engineering record

Source document: TUTOR Frontier Module 17 Agentic Systems Tool Use