## Trigger Use when choosing, scoping, or reviewing a research/development project in the AI era, especially when the project might become "just a prompt", "just a wrapper", or a workflow that a frontier agent will soon perform directly. ## Evidence Basis - Capability: Stanford AI Index 2026 reports accelerating frontier capabilities, SWE-bench Verified nearing saturation, OSWorld agent task success rising to about 66%, and expanding AI-for-science coverage. - Software: Bain's 2025 SaaS analysis frames agentic AI as automating tasks and replicating workflows; high-risk workflows lack proprietary data depth, switching/network friction, standards control, and regulatory/certification barriers. - Platforms: MCP and agent-to-agent protocols make tools and data easier for general agents to access, so "we integrated X" is not enough unless the integration exposes an owned substrate. - Science: AI Scientist, AI Scientist-v2, AlphaEvolve, and agentic science work show ideation, experiment execution, manuscript drafting, algorithm discovery, and evaluator-driven loops becoming automatable. - Theory: Sutton's bitter lesson says scalable general methods tend to overtake hand-built special cases; complementary-assets strategy says value capture moves to scarce assets around a commoditizing technology. ## Fix - First assumption: the model layer is a rising external capability, not the moat. Do not bet on today's model weakness, missing tool use, context limit, price, or latency as the reason the project will matter. - Run the substitution test: "If a frontier agent with browsing, code execution, APIs, memory, and cheap inference can do 80% of this workflow, what value remains here?" - Identify at least one durable substrate: - Earned data: proprietary logs, annotations, failure cases, negative results, lab records, user traces, or domain corpora with rights and provenance. - Workflow ownership: approvals, handoffs, incentives, procurement, review norms, community, distribution, or customer relationships. - Verification loop: gold sets, tests, simulators, benchmarks, physical assays, peer review, audit trails, or outcome metrics that can reject bad model output. - Execution control plane: secure write-back, permissions, APIs, MCP/agent surfaces, observability, rollback, and human override around systems the project owns. - Physical/institutional position: instruments, wet labs, clinical access, regulated processes, supply chain, field operations, or trusted institutional role. - Convert model progress into a dividend: - Keep model providers swappable behind thin adapters. - Store intermediate artifacts as structured data, not only chat transcripts. - Expose agent-readable tools and schemas, not only human UI. - Let models propose; let owned evaluators, tests, users, or physical reality select. - Make every model upgrade reduce cost, expand coverage, improve recall, or increase throughput without rewriting the core project. - Prefer project shapes where general agents help rather than replace: - Data flywheels that get better through real use. - Evaluator/orchestrator systems where search is cheap and verification is scarce. - Research infrastructure that connects models to unique data, instruments, simulations, and reproducibility checks. - Agent-native workflow layers around owned organizations, not generic chat frontends. - Trust, provenance, safety, redaction, compliance, and audit layers that become more necessary as agents gain autonomy. - Reject or re-scope when the value is mostly: - A generic LLM call over generic public input. - A UI around another company's model with no owned data, workflow, distribution, or evaluator. - A hand-coded expert system likely to be absorbed by stronger search/learning. - A benchmark demo with no path to real write-back, adoption, or measured outcome. ## Project Review Template Fill this before starting a serious project: | Row | Required answer | | --- | --- | | Substitution risk | What can the next frontier agent replace? | | Durable substrate | What scarce asset remains outside the model layer? | | Agent interface | How will stronger agents use or extend this system? | | Model-upgrade dividend | What improves automatically as models improve? | | Verification loop | What rejects false, unsafe, shallow, or non-novel output? | | Ownership moat | What data/workflow/trust/institution/distribution is hard to copy? | | Decision | Green / yellow / red; continue, re-scope, or kill. | Decision rule: - **Green**: durable substrate + agent interface + model-upgrade dividend + verification loop are all credible. - **Yellow**: one column is weak; prototype only after naming the missing substrate or evaluator. - **Red**: the project depends mainly on current model limitations or generic prompt/UI assembly; do not build as-is. ## Sources - Stanford HAI, "The 2026 AI Index Report": https://hai.stanford.edu/ai-index/2026-ai-index-report - Rich Sutton, "The Bitter Lesson": https://bitterlesson.ai/ - Bain & Company, "Will Agentic AI Disrupt SaaS?": https://www.bain.com/insights/will-agentic-ai-disrupt-saas-technology-report-2025/ - Anthropic, "Introducing the Model Context Protocol": https://www.anthropic.com/news/model-context-protocol - Xin, Kitchin, and Kulik, "Towards agentic science for advancing scientific discovery": https://www.nature.com/articles/s42256-025-01110-x - Lu et al., "The AI Scientist": https://arxiv.org/abs/2408.06292 - Yamada et al., "The AI Scientist-v2": https://arxiv.org/abs/2504.08066 - Google DeepMind, "AlphaEvolve": https://deepmind.google/blog/alphaevolve-a-gemini-powered-coding-agent-for-designing-advanced-algorithms/ ## Reuse Rule - **Load when**: evaluating whether a research/development idea will remain useful as frontier models and general agents get stronger; or when turning a model-wrapper idea into a durable system. - **Do not load when**: the task is a narrow implementation detail where the project strategy is already fixed and no AI-era substitution risk is being considered.