Agents — Roadmap

How LLMs go from answering questions to taking actions, using tools, and coordinating with each other.

1. Agent Fundamentals

• What makes something an "agent" vs a single LLM call
• Tool use / function calling — how models decide to call tools and parse results
• Designing tools: naming, descriptions, argument schemas
• Agent memory: short-term (context) vs long-term (persisted) memory

2. Context Engineering

• Context engineering vs. prompt engineering, and why it's its own discipline (context rot)
• The write/select/compress/isolate framework
• Compaction, structured note-taking, just-in-time retrieval, sub-agent isolation

3. Agent Architectures

• ReAct (Reasoning + Acting) loop
• Plan-and-Execute pattern
• Reflection / self-critique loops
• Routing and the supervisor pattern
• Single-agent vs multi-agent systems
• Sub-agents and orchestrator patterns
• Memory and state: short-term vs long-term, episodic vs semantic
• Human-in-the-loop: approval gates, escalation, active correction
• Durable execution: retries, checkpoints, long-running agents
• Sandboxing for code/computer-use actions
• Application patterns: browser agents, coding agents, computer-use agents, autonomous research agents

4. Protocols

• MCP (Model Context Protocol) — standardizing how agents connect to tools and data sources
  • MCP servers vs clients
  • Resources, tools, and prompts in MCP
  • Protocol Deep Dive — transport (stdio/HTTP+SSE), sessions & lifecycle, tool discovery & JSON Schema, errors, auth/authz, security, enterprise deployment, building a server
• A2A (Agent-to-Agent) — how independent agents discover and communicate with each other
  • Agent cards / capability discovery
  • Task delegation between agents

5. Multi-Agent Systems

• Coordination patterns: hierarchical, peer-to-peer, blackboard
• Shared state and conflict resolution
• When multi-agent actually beats a single well-prompted agent (and when it doesn't)

6. Common Problems & SOTA Solutions

• Agent loses context over long tasks → memory systems, summarization, external state stores
• Agent gets stuck in loops → step limits, reflection checkpoints, human-in-the-loop escalation
• Tool-calling hallucination (calling tools that don't exist, malformed args) → strict schemas, validation, retries
• Runaway cost/latency in multi-agent chains → caching, cheaper models for sub-tasks, parallelization