seven targets · one architecture recommendation · evidence-based

Platform Options

Seven platforms, one architecture recommendation. The Architecture Agent scores each platform against the PRD — evidence-based, no vendor default.

How the recommendation is built: the PRD goes to the Architecture Agent, which scores every platform against its weighted rubric, ranks them with rationale and watch-outs, and hands the choice to the user — who is always the final decider before the build pipeline runs.

Enterprise platformsManaged knowledge & work systems

Glean

Enterprise knowledge retrieval — connector-based data access, workflow automation inside the Glean ecosystem.

Output · agent-config.json + architecture guide + deploy instructions + test scenarios
Constraint · a Glean connector per data source; agents are stateless, no persistent caching.

Monday.com

Work OS — boards, automations, AI agents, 24+ column types, native integrations, and MCP programmability.

Output · board schemas + automation configs + GraphQL mutations + agent JSON
Constraint · board-first data model; rate limits by plan tier (1k–25k calls/day).

API / agent platformsModel APIs & managed agents

Claude API

Maximum reasoning capability — complex analysis, tool use, multi-step agentic workflows, extended thinking, full parameter control via the Anthropic SDK.

Output · system prompt + tool definitions + orchestration code + evaluation harness
Constraint · API key required; no built-in UI — you bring the interface.

OpenAI Assistants

Built-in code interpreter, file search (RAG), and function calling; managed threads and persistent conversations.

Output · assistant config + function schemas + thread management code
Constraint · locked into the OpenAI ecosystem; usage-based pricing scales with complexity.

Generic / platform-agnostic

A portable system prompt — platform TBD, simple Q&A, or a proof of concept. Build the agent first, decide where it runs later.

Output · standalone system prompt + evaluation criteria + integration guide
Constraint · no deployment target; the prompt is portable, implementation is on you.

Deployment targetsWhere the agent runs

Full-stack cloud

Production-grade MVP stack: Next.js on Vercel, FastAPI on Fly.io, Supabase for auth and database. Dual LLM support — DeepSeek for prototyping, Anthropic for demos. A live, shareable, credible URL.

Output · main.py + Next.js app + schema.sql + fly.toml + Dockerfile + six deploy prompts + DEPLOY.md
Constraint · three cloud accounts (Vercel, Fly.io, Supabase); first deploy about 20 minutes.

Local

A self-contained web app that runs on your machine — FastAPI + SQLite + browser UI. Zero cloud dependencies; ideal for demos and prototypes.

Output · main.py + Jinja2 templates + Dockerfile + docker-compose.yml
Constraint · single machine only; runs entirely local.

Human handoffNot an agent at all

Developer spec

Hand a full-stack developer the spec and let them build it. The PRD goes in; a decision-complete spec comes out; the developer owns implementation.

Output · BRIEF.md (requirements contract) + SPEC.md (implementation reference)
Constraint · a human builds it; the spec is decision-complete, and the stack, schema, and API are the developer’s call.

The Architecture Agent scores every platform against the PRD using each platform’s weighted rubric. The recommendation is evidence-based — no vendor default, no lock-in. Platforms rank 1–7 with rationale, best-fit signals, and watch-outs. The user always has the final choice; select one and the build pipeline takes over.