Bespoke
Agentics

Intelligent by Design

Explore

The gap between idea and implementation
is closing faster than anyone expected.

AI agents that can reason about full project context, decompose complex requirements, generate production code in sandboxed environments, and push changes to repositories autonomously are no longer theoretical. They're being built right now. The question isn't whether this transformation happens — it's what the delivery platform looks like when it does.

Bespoke Agentics exists to explore that question through building. Not through decks or demos, but through working software that structures the messy reality of professional delivery — requirements, risks, workstreams, design context, source control — and gives AI agents the structured knowledge they need to reason about real projects.

Delivery knowledge is the moat

Most AI coding tools treat code generation as an isolated act. Real delivery requires structured context: which requirements map to which workstreams, what the design constraints are, which repos own which capabilities. The platform that structures this knowledge wins.

Agents need guardrails, not just prompts

Autonomous code generation at scale demands sandboxed execution, human review queues, audit trails, and verification pipelines. The hard problem isn't making AI write code — it's making AI write code you can trust in production.

An agentic delivery platform,
open-sourced in alpha.

Foundry is a multi-tenant agentic delivery platform that transforms how agencies and software delivery businesses turn plans, conversations, and requirements into working deliverables. Feed it plans and conversations — Foundry decomposes them into structured requirements, reasons about implementation, provisions AI sandboxes, and ships code to repos.

This is a solo founder build by Quintin Henry. It's being open-sourced not because it's finished, but because the problem it's attacking — the full idea-to-implementation loop — is too important and too large for closed development. The software is in an alpha state. The architecture is real, the codebase is substantial, and the system works end-to-end. What it needs now is scrutiny, contribution, and the compounding insight that comes from building in public.

Foundry is engagement-type agnostic. It works for platform migrations, greenfield builds, system integrations, and ongoing product development. The reference client is Burlington Medical — 118 requirements, 8 skills, 7 workstreams — running through the full delivery pipeline.

Stage Alpha · v0.1.0
Frontend Next.js 16 + React 19
Backend Convex (55+ tables)
AI Claude (Opus / Sonnet)
Auth Clerk (multi-tenant)
Sandbox Cloudflare Workers + Docker
Desktop Tauri 2 (Rust + Vite)
License Open Source
55+
Database Tables
405
UI Components
52
App Routes
127K
Lines Added

Four-process distributed system.

A browser/desktop client connected via WebSocket to a reactive backend, with AI inference and sandboxed execution handled by edge workers. Every query is indexed, every mutation is auth-gated, every agent action is auditable.

Browser + Desktop

Next.js 16 & Tauri

52 routes, thin page wrappers over shared @foundry/ui components. Real-time subscriptions via Convex WebSocket. Desktop app shares the same component library through Tauri 2.

Backend

Convex Cloud

55+ tables across 6 functional domains. Server functions for queries, mutations, and actions. AI actions assemble context and call Claude. Webhook handlers for GitHub, Atlassian, and Clerk.

AI Inference

Agent Worker

Hono + Anthropic SDK on Cloudflare Workers. Structured output endpoints for requirement analysis, task decomposition, and codebase analysis. Three-tier model deployment.

Sandbox

Execution Engine

Durable Objects orchestrating Docker containers. 10-stage provisioning state machine. Ephemeral Claude Code environments for autonomous code generation, verification, and auto-commit to repos.

Auth

Clerk Multi-Tenant

Organizations as tenants. JWT validation on every request. Row-level security via assertOrgAccess() on every query and mutation. Webhook-synced user records.

Integrations

GitHub & Atlassian

GitHub App for repos, PRs, and webhooks. Atlassian OAuth for Jira and Confluence. Durable event buffer pattern with exponential backoff retry.

What works, what's next.

This is alpha software. The core delivery pipeline is stable and running end-to-end. The dominant development theme over the past two weeks has been deepening AI integration and building the observability layer for agent-driven delivery.

Shipped & Stable

  • Core delivery pipeline — requirements, skills, tasks, workstreams
  • Sandbox execution system with 10-stage provisioning
  • Agent Activity dashboard with full audit trail
  • Design context pipeline with AI vision analysis
  • Repository picker across tasks and workstreams
  • Task verification pipeline
  • Google Drive import source
  • Service resilience layer (auto-reconnect, health monitoring)
  • Billing system (3 tiers)
  • Biome lint + format enforcement with hooks
  • GitHub App + Atlassian integrations
  • Clerk multi-tenant auth with row-level security

In Progress

  • Codebase analysis — AI-powered requirement-to-implementation mapping with human review queue
  • Semantic code search — vector embeddings and cosine similarity for requirement analysis
  • Test coverage initiative — 28% to 90% via 4-agent parallel strategy
  • Orchestrator decomposition — extracting the 4,142-line state machine into focused modules

Building in public because the
problem demands it.

The idea-to-implementation gap is one of the most consequential problems in software right now. Closing it requires combining structured delivery knowledge, AI reasoning, sandboxed execution, and human oversight in ways nobody has fully figured out yet. That kind of problem benefits from being open.

Scrutiny sharpens architecture

A 55-table schema, a 4,142-line orchestrator, and a multi-process distributed system all benefit from external eyes. The codebase is real and substantial. It needs critique as much as contribution.

The loop is bigger than one team

Requirements intake, AI decomposition, sandbox execution, verification, PR generation, design context cascading — each piece is a deep problem. Opening the platform invites domain expertise at every layer.

Reference implementation

Whatever the future of agentic delivery looks like, there's value in having a working, documented, end-to-end system that people can study, fork, and build on. Foundry is that system, warts and all.