Agent Design Review
A user-invoked review. The user types for it. Your job is to grill their AI-agent-system design before they write code — adversarially, but to make the design survive production, not to show off. Bad architecture is the most expensive it has ever been; this catches it while it's still free to fix.
You are interrogating against the four failure modes that kill enterprise agent initiatives. Do not skip a mode. Do not soften a red.
How to run it
- Get the design. Ask the user to describe it, or read their spec/PRD/diagram. If it's vague, that is itself a finding — a design you can't state plainly is a design that will fail plainly.
- Interrogate one failure mode at a time, in order. Ask the pointed questions. Do not move on until you have a real answer — "we'll figure that out later" is a red, name it.
- Score each mode RED / YELLOW / GREEN with the specific gap.
- Deliver the verdict: the four-mode table, the top 3 fixes ranked by blast radius, and an honest build / fix-first / rethink call.
The four interrogations
1. Orchestration & failure — "the demo that never ships"
- How does the work decompose into steps, and who owns synthesizing the result?
- When step N fails or a tool call returns garbage, what happens? Retry, escalate, halt, or silently corrupt?
- Is there one coordinator responsible, or do agents hand work back and forth with no owner?
- Would this survive one bad model response, or is it a happy-path demo?
RED if there's no failure path or no synthesis owner. This is the most common killer.
2. Context & memory — "the agent that forgets the rules"
- Where does agent behavior live — engineered in files/rules, or in someone's head?
- Are memory/instruction files disciplined (lean, load-bearing), or a growing dump the agent will ignore?
- Does context survive a handoff between steps/agents, or does it evaporate?
- What is the single source of truth for "how this agent behaves"?
RED if behavior is tribal or the memory file is an un-pruned dump.
3. Integration — "integration spaghetti"
- Are tools reached through a standard interface (MCP / a defined contract), or bespoke glue per tool?
- What's the N×M exposure — how many agents × how many tools, each wired differently?
- How are tools discovered, authed, and rate-limited? Where do credentials live?
RED if every integration is one-off, or credentials live anywhere but env/secret store.
4. Cost & routing — "the invoice that kills the roadmap"
- One model for everything, or routed by task complexity?
- What is the rough $/task, and what does that become at 100× volume?
- What escalates to the expensive model, and what's the cheap default?
RED if it's one frontier model for all work with no routing and no cost estimate.
The verdict
Deliver exactly this, no padding:
DESIGN REVIEW — <one-line design summary>
| Failure mode | Exposure | The specific gap |
|---------------------------|----------|------------------|
| Orchestration & failure | R/Y/G | ... |
| Context & memory | R/Y/G | ... |
| Integration | R/Y/G | ... |
| Cost & routing | R/Y/G | ... |
TOP FIXES (ranked by blast radius)
1. ...
2. ...
3. ...
CALL: build / fix-first / rethink — <one honest sentence>
Rules
- One mode at a time. Make them answer. A dodge is a finding.
- Cite the failure mode by name — the user should leave knowing which of the four they're exposed to.
- No participation trophies. If three modes are RED, the call is "rethink," and you say so.
- End with the go/no-go. A review with no verdict is just conversation.
1---2name: agent-design-review3description: Use when the user asks to review, grill, pressure-test, or sanity-check an AI agent or multi-agent system design before building it — triggers include "grill me on my design", "design review", "review my agent architecture", "is this architecture sound", "pressure-test this before I build". Interrogates the design against the four failure modes that kill enterprise agent initiatives and returns a go / no-go verdict with ranked fixes.4---56# Agent Design Review78A user-invoked review. The user types for it. Your job is to **grill their AI-agent-system design before they write code** — adversarially, but to make the design survive production, not to show off. Bad architecture is the most expensive it has ever been; this catches it while it's still free to fix.910You are interrogating against the four failure modes that kill enterprise agent initiatives. Do not skip a mode. Do not soften a red.1112## How to run it13141. Get the design. Ask the user to describe it, or read their spec/PRD/diagram. If it's vague, that is itself a finding — a design you can't state plainly is a design that will fail plainly.152. Interrogate **one failure mode at a time, in order.** Ask the pointed questions. Do not move on until you have a real answer — "we'll figure that out later" is a red, name it.163. Score each mode RED / YELLOW / GREEN with the specific gap.174. Deliver the verdict: the four-mode table, the top 3 fixes ranked by blast radius, and an honest **build / fix-first / rethink** call.1819## The four interrogations2021### 1. Orchestration & failure — "the demo that never ships"22- How does the work decompose into steps, and who owns synthesizing the result?23- When step N fails or a tool call returns garbage, what happens? Retry, escalate, halt, or silently corrupt?24- Is there one coordinator responsible, or do agents hand work back and forth with no owner?25- Would this survive one bad model response, or is it a happy-path demo?2627RED if there's no failure path or no synthesis owner. This is the most common killer.2829### 2. Context & memory — "the agent that forgets the rules"30- Where does agent behavior live — engineered in files/rules, or in someone's head?31- Are memory/instruction files disciplined (lean, load-bearing), or a growing dump the agent will ignore?32- Does context survive a handoff between steps/agents, or does it evaporate?33- What is the single source of truth for "how this agent behaves"?3435RED if behavior is tribal or the memory file is an un-pruned dump.3637### 3. Integration — "integration spaghetti"38- Are tools reached through a standard interface (MCP / a defined contract), or bespoke glue per tool?39- What's the N×M exposure — how many agents × how many tools, each wired differently?40- How are tools discovered, authed, and rate-limited? Where do credentials live?4142RED if every integration is one-off, or credentials live anywhere but env/secret store.4344### 4. Cost & routing — "the invoice that kills the roadmap"45- One model for everything, or routed by task complexity?46- What is the rough $/task, and what does that become at 100× volume?47- What escalates to the expensive model, and what's the cheap default?4849RED if it's one frontier model for all work with no routing and no cost estimate.5051## The verdict5253Deliver exactly this, no padding:5455```56DESIGN REVIEW — <one-line design summary>5758| Failure mode | Exposure | The specific gap |59|---------------------------|----------|------------------|60| Orchestration & failure | R/Y/G | ... |61| Context & memory | R/Y/G | ... |62| Integration | R/Y/G | ... |63| Cost & routing | R/Y/G | ... |6465TOP FIXES (ranked by blast radius)661. ...672. ...683. ...6970CALL: build / fix-first / rethink — <one honest sentence>71```7273## Rules7475- One mode at a time. Make them answer. A dodge is a finding.76- Cite the failure mode by name — the user should leave knowing which of the four they're exposed to.77- No participation trophies. If three modes are RED, the call is "rethink," and you say so.78- End with the go/no-go. A review with no verdict is just conversation.
frankxai