에이전트 시스템 키트

Multi-agent design: roster, topology, handoff contracts, shared memory, A2A and MCP links, budgets, cascading-failure controls, oversight, OWASP ASI mapping

Specify a tool-using AI agent: autonomy level, tools and permissions, guardrails, security, trajectory evals, cost

Agent tool permissions: tool inventory, capability classes, least agency, non-human identity, approval tiers, sandboxing, limits, audit, revocation, tests

Human oversight: objectives, decision points (in, on, over the loop), approval gates, override and stop, overseer training, automation bias, UI, logs

A2A agent spec: Agent Card fields, skills, bindings, security schemes, operations, task states, parts, push, errors, discovery, example card, conformance

Specify an MCP server (spec revision 2026-07-28): transport, tools, resources, prompts, OAuth, security, testing

System prompt spec: purpose, model parameters, role, instructions, safety rules, output schema, examples, tools, tests

Repository instructions for AI coding agents: setup, build and test commands, conventions, architecture map, PR and security rules

Agent trust boundaries, assets, autonomy, ASI01-ASI10 threat register, MITRE ATLAS mapping, L x I scoring, mitigations, residual risk, test cases

Policy taxonomy, input, output and tool-call guards, thresholds and actions, false-positive and latency budgets, guard evaluation, OWASP LLM 2026 mapping

Red team scope, threat model, NIST AI 100-2 classes, OWASP LLM 2026 and ASI01-ASI10 coverage, MITRE ATLAS, test cases, severity, findings, retest

Agent task suite, success, trajectory, tool-call and argument metrics, sandboxed harness, online tests, calibrated graders, statistics, ASI safety cases, gates

LLM app evaluation: test sets and red team, metrics, LLM-as-judge calibration, human review, CIs, release gates

LLM/agent telemetry: objectives, span model, gen_ai.* attributes, token/latency/cost metrics, prompt capture controls, eval hooks, dashboards, SLOs, retention

Launch scope and tier, 18 AI gates with owners, eval results vs thresholds, red team findings, Article 50 and register checks, exceptions, go/no-go, post-launch

Report an AI incident or near miss: harm, severity, timeline, root cause, containment, notifications incl. AI Act Art. 73