Agents, MCP and skills

Trace a model-requested tool call through schema validation, execution and a returned observation.

1. A1.1Tools as contracts
2. A1.2JSON schemas
3. A1.3The tool-call loop
4. A1.4Errors and observations
5. A1.5Structured outputs
6. A1.6Side effects and identity

Design a bounded observe-act loop with explicit state, stopping conditions and a verifiable task outcome.

1. A2.1State and actions
2. A2.2Planning
3. A2.3Observe and revise
4. A2.4Memory and context
5. A2.5Stopping and budgets
6. A2.6Reflection and evaluation

Explain MCP roles and lifecycle and trace a tool-discovery and tool-call exchange under a pinned protocol version.

1. A3.1Hosts, clients and servers
2. A3.2Lifecycle and capabilities
3. A3.3Transports
4. A3.4Tools
5. A3.5Resources and prompts
6. A3.6A Python server

Write a scoped skill package and manage its instructions and references within a task's context budget.

1. A4.1Instructions and data
2. A4.2A skill directory
3. A4.3Progressive disclosure
4. A4.4Context windows
5. A4.5Compaction and memory
6. A4.6Skill evaluation

Choose when to delegate independent work and define handoffs, shared-state rules and evaluation.

1. A5.1Task decomposition
2. A5.2Roles and contracts
3. A5.3Parallel work
4. A5.4Handoffs
5. A5.5Orchestration and tracing
6. A5.6Evaluation

Identify instruction-injection attempts and design permission, sandbox, approval and evaluation boundaries for an agent.

1. A6.1Prompt injection
2. A6.2Least privilege
3. A6.3Sandboxing
4. A6.4Human review
5. A6.5Evaluations
6. A6.6Observability and incidents