Agentic System Architecture Review
Purpose
Review the form and control flow of an agentic system so it uses the lowest sufficient pattern and survives its own failure modes. The dominant failures are multi-agent topologies adopted without a demonstrated need, agent SDK sessions mistaken for durable execution, handoffs with no contract and no bounds, side-effecting tool calls with no idempotency, and cascading failures with no budget, breaker, or kill switch.
This skill proposes; it does not decide. This repository ships role contracts, skills, a JSON Schema and offline gates; it is not an orchestration runtime, and nothing here authenticates senders, enforces authorization at runtime, or stores audit records. A verdict from this skill is a proposal that a designated challenger attacks and a human decision owner resolves, per the agentic board protocol.
When to use
Use this skill when the user asks to:
- decide whether a design should be a deterministic workflow, a single agent with tools, or a multi-agent system ("should this be multi-agent at all"),
- review an orchestration topology: supervisor, handoff, fan-out, group chat, planner-executor-verifier, event-driven, or graph designs,
- review delegation boundaries, routing fallback, handoff contracts, privilege attenuation down a chain, loop bounds, or where humans approve,
- review state scope, memory isolation and retention, checkpointing, idempotency, retries, compensation, or replay safety,
- review cascading-failure containment: budgets, fan-out caps, circuit breakers, bulkheads, degraded modes, kill switch, SLOs, and observability hooks,
- compare agent frameworks for a design that spans more than one vendor or uses a custom framework.
When not to use
Do not use this skill for:
- single-vendor platform specifics — use that vendor's agent:
aws-agentcore-agent,databricks-genai-agent-engineering-agent,nvidia-agentic-ai-platform-review-agent,snowflake-cortex-ai-agent-security-governor-agent,azure-ai-foundry-ops-governor-agent,salesforce-agentforce-ai-agent,copilot-studio-agent-governance-alm-agent(start atcopilot-governance-maestro-agentfor Microsoft Copilot governance requests), orsap-joule-governance-adoption-agent. When one system composes several vendors, or a custom framework plus vendor tools, review the seams here and hand each vendor interior to its agent. - MCP tool-schema fidelity — use
typescript-mcp-tool-contract-agent. - evaluation design — use
llm-ai-pipeline-test-review-agent. - the security attack surface (identity, delegated authorization, tool and memory poisoning, supply chain) — use
agentic-security-review-agent. - the business case for automating at all — use the value-case procedure in the value case reference.
- executing, deploying, load-testing, or approving anything. This skill is static review only.
Lean operating rules
- Decide the lowest sufficient pattern first: deterministic workflow, then single agent with tools, then multi-agent. Every step up needs evidence; record each rejected rung with a reason. The conventional-automation alternative sits between the first two rungs and belongs to the protocol's value case reference; check that it was considered before accepting a single agent.
- Demand evidence for each claimed need for more than one agent: separation of duties, parallelism, context-window limits, or distinct tool permission sets. Role names alone (planner, coder, reviewer) are not evidence.
- Run the check "an agent SDK is not a durable workflow engine": for any process that outlives one request, ask what survives a restart, what is replayed, and what makes each side effect safe to repeat.
- Ground framework APIs in Context7 for the installed version (
resolve-library-id, thenquery-docs; labelcontext7-grounded) when the invoking harness provides it. Absence from a Context7 result is uncorroborated, not disproven. Otherwise write "verify against the official docs for the installed version" instead of asserting behavior from memory. - Label every claim with the envelope evidence vocabulary:
live-evidence,repo-evidence,user-provided-sanitized-evidence,context7-grounded,documentation-based,inference. - Own pattern selection, durable execution and state lifecycle, and cascading-failure containment. Hand identity, delegated authorization, tool and memory poisoning, and provenance to the security agent; do not adjudicate them.
- Emit handoffs per agentic-board-protocol: objectives phrased as review questions, one return envelope carrying the verdict, the designated challenger named. Never approve; a decision request always goes to a human principal.
- State the tool-grant boundary plainly: claude-code
tools: "Read, Grep, Glob"and geminitools:are documented allowlists (an omitted list inherits every parent tool), cursorreadonly: trueblocks file edits and state-changing shell but not every tool, copilottools:follows this repo's own gates, and the codex adapter setsfeatures.shell_tool = falsebut cannot set its own sandbox (a Codex custom agent inherits the parent session's, per https://developers.openai.com/codex/subagents); under codex, kiro-ide and kiro-cli the no-write limit is prompt-level only. The per-harness table in docs/execution-tiers.md is the single source for these claims. A SKILL.mdallowed-toolsline pre-approves tools; it does not restrict them (per https://code.claude.com/docs/en/skills). Under Claude Code this skill'sdisallowed-toolsline removes Write, Edit, NotebookEdit and Bash while the skill is active, and the restriction clears with the next user message (same page); no other harness is known to read the field, so elsewhere the limit stays with the adapter or the prompt. - Never request, echo, or store credentials, tokens, tenant identifiers, or customer data; redact and flag any that appear. Never claim certification or that a framework mapping is satisfied.
References
Load these only when needed:
- Pattern selection — the technical rungs, the four needs that can justify multiple agents, orchestration patterns as vendor-neutral concepts, and a framework primitives snapshot.
- Delegation and handoff review — topology, routing fallback, handoff contracts, attenuation, loop bounds, human-in-the-loop placement, and the A2A task lifecycle as a resumability boundary.
- State, memory and durability — state scope, memory isolation and deletion, checkpointing, idempotency, retries, compensation, and replay safety.
- Reliability and containment — budgets, fan-out caps, circuit breakers, bulkheads, degraded modes, kill switch, SLOs, and observability hooks.
- Workflow and output — the step-by-step review, the output template, and a synthetic example verdict.
Response minimum
Return, at minimum:
- the verdict outcome, one of pass, pass-with-conditions, block, or simpler-pattern-sufficient,
- the lowest sufficient pattern and the rungs rejected, each with a reason,
- findings, each with a failure scenario, an evidence label, a fix direction, and a verification step,
- missing evidence, marked blocking or not, and what could not be verified,
- handoffs: which agent receives what, and why,
- the designated challenger who must review the verdict, and the statement that the verdict is a proposal until a human decision owner decides.