Architecture Playground

Add components, drag them, connect them (click Connect, then click source and target). The analyzer explains strengths, weaknesses, security and reliability risks, latency, cost — and how to simplify.

Presets
Add
9 components · 8 connections
UserGuardrailAgentMemoryRetrieverVector DBMCP Server: GitHubMCP Server: SlackHuman Approval

Click a connection to remove it. Drag nodes to arrange. Select a node to label or delete it.

Latency
tens of seconds: delegation chains
Cost
context grows per iteration — budget per task
Strengths
  • One agent with a clear owner of the loop — easy to trace and evaluate.
  • MCP servers: integrations are discoverable and reusable across applications instead of hand-coded per app.
  • Retrieval grounds answers in your data instead of model memory.
  • Memory gives session and cross-session continuity.
  • Approval gate present: risky actions are reviewed before execution.
  • Guardrails in the path — remember they are one layer; the tool permission layer must enforce policy in code.
Weaknesses
  • RAG without an evaluator: retrieval quality (recall@k) and faithfulness are unmeasured — the two most common silent failures.
  • Memory needs write rules, expiry and per-user scoping; unbounded memory pollutes context.
Security risks
  • MCP tool output is untrusted content — a repository file or Slack message can carry injected instructions.
  • Persisted memory is personal data: define retention, deletion and who can read it.
Reliability risks
  • No evaluator: regressions are invisible until users complain.
  • Agents need step limits, token/cost budgets and loop detection — none are visible on a diagram; make sure they exist in code.
  • Tool calls fail: classify retryable errors, add timeouts and idempotency keys for writes.
  • Provider outages happen: define a fallback model or graceful degradation.
Possible simplifications
  • No reranker: fine at first. Add one only when evals show precision problems in the top-k.

How the analysis works

Rule-based, transparent, and deliberately opinionated in favor of simplicity.

The analyzer counts model hops, tools, retrieval stages and approval gates to estimate latency and cost; checks for missing guardrails, evaluators and approval gates; flags orphaned components, unowned multi-agent shapes and oversized tool lists; and always asks whether a simpler rung of the ladder would do. It cannot see step limits, budgets or permissions — those live in code, and you should add them.