| name | staged-validation |
| description | Design a phased go/no-go validation roadmap for R&D or platform teams building systems that will be adopted by multiple downstream teams. |
| when-to-use | Use when an R&D group has a working prototype and needs a structured path from proof-of-concept to production-ready platform, with clear success criteria at each stage. |
| principles | ["Bias Towards Action","Governance Hierarchy","Rate of Improvement","First Principles","HITL and Guardrails"] |
Staged Validation Skill
Purpose
Convert a prototype or incomplete architecture into a sequenced validation roadmap. Each phase has a focused scope, explicit success criteria, and a go/no-go gate. Prevents R&D teams from scaling prematurely or building breadth before depth.
Agent Instructions
Step 1: Identify the Core Value Claim
Before designing phases, extract the system's primary value claim — the one thing it must do reliably for adoption to make sense.
Ask:
- What is the single agent or mechanism most novel and differentiated in this system?
- If only one component worked end-to-end against a real target, which delivers the most value?
- What failure mode would make downstream teams abandon adoption?
The component that answers all three questions is Phase 1.
Pareto principle: Validate the 20% of the system that delivers 80% of the value before building the remaining 80%.
Step 2: Design the Phase Structure
Each phase follows this pattern:
Phase N
Focus: [single agent or capability pair]
Scope: [what is real vs. still simulated]
Gate: [measurable success criteria — numbers, not feelings]
Output: [what artifact or signal proves the gate was met]
Effects: [top second-order effects expected at this phase boundary — what does passing this gate trigger in adjacent systems or teams?]
Decision: [go / no-go / iterate]
Standard phase sequence for agentic systems:
| Phase | Typical Focus | Gate Example |
|---|
| 1 | Core agent — real input, real output | Agent produces valid output for ≥3 real targets with <N retries avg |
| 2 | Orchestrator + core agent — end-to-end | Natural language prompt → stored data in <X minutes reliably |
| 3 | Add resilience layer (monitoring/healing) | Measurable reduction in failure rate vs. baseline |
| 4 | Add validation layer (QA/audit) | Output accuracy meets defined threshold; feedback loop active |
Adapt phase count and focus to the specific system. Never add a new phase before the previous gate is met.
Step 3: Define Platform Foundations
For systems that will be consumed by multiple teams, identify the platform foundations that must be hardened early — regardless of which phase the team is in:
| Foundation | Why It Can't Wait |
|---|
| Durable state | Teams cannot adopt a system whose task state is lost on restart |
| Agent contracts | Teams cannot integrate against agents with undocumented or unstable APIs |
| Versioned prompts | Teams cannot trust a system whose behavior changes without notice |
| Basic HITL gate | Teams cannot deploy a system that can't be stopped or overridden |
These are not Phase 4 concerns. They are Phase 1 prerequisites for platform readiness.
Step 4: Define the Feedback Loop Per Phase
Each phase must produce a learning signal — not just a pass/fail:
| Signal Type | Example |
|---|
| Performance metric | Avg LLM retries per job, success rate per target |
| Quality metric | QA pass rate, field mismatch count |
| Cost metric | LLM calls per task, tokens per decision |
| Reliability metric | Crash recovery rate, uptime |
Instrument the signal in Phase 1. Don't retrofit observability.
Step 5: Advise on Depth vs. Breadth
R&D teams building platforms are often tempted to demonstrate breadth (all agents working at once) over depth (one agent working perfectly).
Advise depth first when:
- The system is being presented for adoption by other teams
- Any component is still simulated
- Success criteria are not yet measured with real data
- The team cannot answer: "What is our baseline failure rate?"
Breadth becomes appropriate when:
- Phase 1 gate has been met and documented
- At least one downstream team has integrated Phase 1
- The team can demonstrate a real feedback loop (not a demo loop)
Step 6: Write the Advisory Summary
Produce a summary suitable for sharing with the R&D team and their stakeholders:
- Reframe statement — what the team is actually building (platform vs. demo)
- Phase table — scope, gate, output, and effects for each phase
- Platform foundations — what to harden regardless of phase
- Depth vs. breadth guidance — one-line principle the team can repeat
- Second-order effects at scale — when this platform reaches full downstream adoption, what does it change in the teams consuming it? Identify: any behaviors or incentives that shift, any adjacent systems that must change, and any feedback loops that activate. Flag any effect that would require design changes if discovered after Phase 3.
- Immediate next actions — 3–5 specific, executable items for their next sprint
Output Format
- Core value claim (one sentence)
- Phase roadmap table (Phase / Focus / Gate / Output / Decision)
- Platform foundations checklist
- Feedback loop design per phase
- Advisory summary (shareable with team and stakeholders)