Knowledge Base / Agentic AI
Agentic AI
An AI agent is a language model running in a loop: it decides what to do, calls a tool, reads the result and continues. That autonomy is what makes agents useful, and it is also what makes them harder to test, secure and audit than ordinary software. Everything an agent can do is something you gave it access to.
This section starts with what an agent actually is, then covers the practical engineering topics in a sensible order: connecting tools safely (including the Model Context Protocol), designing multi-agent systems with clear trust boundaries, building human approvals that people really read, giving agents proper identities and permissions, and measuring and logging their behavior.
Suggested reading order
- 1What Is an AI Agent? Architecture of Planner, Tools, Memory and the Control Loop
A clear explanation of how AI agents differ from chatbots and fixed workflows, the components every agent has, and when an agent is the wrong choice.
- 2Building Secure Agentic AI Workflows: Tools, Permissions and Memory
Security design for AI agents that call tools and APIs: capability scoping, sandboxing, approval gates, vector store isolation, audit logging and runaway-cost controls.
- 3Model Context Protocol (MCP) Security: Risks of Connecting Agents to Tools
What MCP is, the main attack classes against MCP servers and clients (tool poisoning, rug pulls, confused deputy, token passthrough), and a checklist for vetting and running servers safely.
- 4Human-in-the-Loop Design for AI Agents: Approvals That People Actually Read
How to decide which agent actions need approval, how to design approval prompts that prevent mistakes instead of causing click fatigue, and how to measure whether oversight works.
- 5AI Agent Identity and Authorization: Credentials, Delegation and Least Privilege
How to give agents their own identities, use short-lived scoped tokens, delegate on behalf of users safely, and enforce permissions in the tools rather than in prompts.
- 6Evaluating and Testing AI Agents: From Test Suites to Regression Gates
How to test systems whose behavior varies from run to run: building task datasets, scoring outcomes and tool use, handling non-determinism, and running evaluations in CI.
All Agentic AI articles (7)
- 4 min readAI Agent Identity and Authorization: Credentials, Delegation and Least PrivilegeHow to give agents their own identities, use short-lived scoped tokens, delegate on behalf of users safely, and enforce permissions in the tools rather than in prompts.
- 4 min readEvaluating and Testing AI Agents: From Test Suites to Regression GatesHow to test systems whose behavior varies from run to run: building task datasets, scoring outcomes and tool use, handling non-determinism, and running evaluations in CI.
- 3 min readHuman-in-the-Loop Design for AI Agents: Approvals That People Actually ReadHow to decide which agent actions need approval, how to design approval prompts that prevent mistakes instead of causing click fatigue, and how to measure whether oversight works.
- 4 min readModel Context Protocol (MCP) Security: Risks of Connecting Agents to ToolsWhat MCP is, the main attack classes against MCP servers and clients (tool poisoning, rug pulls, confused deputy, token passthrough), and a checklist for vetting and running servers safely.
- 3 min readMulti-Agent Systems: Patterns, Failure Modes and Trust BoundariesCommon multi-agent architectures (orchestrator-worker, pipeline, debate), why they fail in practice, and how to set trust boundaries so one compromised agent cannot take over the rest.
- 3 min readObservability and Audit Logging for AI AgentsWhat to capture in agent traces, how to correlate steps, redact sensitive data, set retention, and use logs for debugging, security investigations and compliance evidence.
- 3 min readWhat Is an AI Agent? Architecture of Planner, Tools, Memory and the Control LoopA clear explanation of how AI agents differ from chatbots and fixed workflows, the components every agent has, and when an agent is the wrong choice.