Skip to content

bitcomplete/agentic-craft

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

186 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

Agentic Craft

Agentic Craft is a reference guide for designing agentic product interfaces beyond chat. It documents interaction patterns for agent orchestration, configuration, tool use, planning, rich action previews, approvals, memory, feedback, multi-agent coordination, monitoring, and observability.

The project is intentionally broader than a UI kit and broader than a chat starter. Components are included so the guidance can be explored, copied, and tested, but the primary artifact is a set of opinionated interface patterns for high-trust agent experiences.

What Is Inside

  • A Next.js reference site with interactive examples for common agent workflows across chat, dashboards, inboxes, task lists, approvals, and monitoring views.
  • A shadcn-compatible registry for reusable agentic interface primitives.
  • Pattern pages that explain behavior, states, rationale, and interaction details instead of only showing static component samples.
  • Research-driven taxonomy work that treats chat as one interface model among many: orchestration, create/configure, monitor, govern, recover, and improve.
  • Workflow templates that show how patterns combine into end-to-end product flows such as review, approval, clarification, memory review, run monitoring, multi-agent handoff, and agent settings.

Reference Areas

  • Conversation: message surfaces, citations, observable work, clarifying questions, and composer behavior.
  • Sources & Artifacts: citation systems, source previews, source-backed documents, and usage meters.
  • Agent Actions: tool calls, subagents, parallel execution, plans, decision flows, clarifying interruptions, and approval gates.
  • Trust & Control Plane: autonomy settings, consent, context scope, provenance, audit trails, cost transparency, and kill-switch behavior.
  • Memory: memory panels, ledger items, provenance previews, memory CRUD, auto-memory, context rings, and privacy controls.
  • Multi-Agent: agent identity, handoff, routing, parallel execution, and inter-agent communication.
  • Feedback: thumbs, corrections, ratings, error reports, escalation, and feedback history.
  • Observability: activity timelines, token usage, session timelines, and error states.
  • Templates: complete workflow references built from the primitives.
  • Registry: local shadcn registry installation, primitives, blocks, and quality gates.

Development

See AGENTS.md for stack details, all commands, the registry invariants, audit gates, and the QA bar. It is the canonical guide for working on the repo, including with AI coding agents. CLAUDE.md adds Claude-Code-specific notes. For design and product rationale see DESIGN.md; for the canonical research behind the patterns see docs/research.md.

pnpm install
pnpm run dev

pnpm is the supported package manager (see packageManager in package.json).

By default, the app runs at:

http://localhost:3000

Registry

Build the shadcn registry artifacts with:

pnpm run registry:build

The source registry lives in registry.json and is designed for GitHub distribution:

npx shadcn@latest list bitcomplete/agentic-craft
npx shadcn@latest add bitcomplete/agentic-craft/review-workflow

The generated JSON registry is served locally from:

http://localhost:3000/r/registry.json

Current registry items:

  • composer
  • tool-call
  • tool-tree
  • reference-item
  • observable-work
  • decision-surface
  • file-lifecycle
  • agent-status-table
  • clarifying-questions
  • action-preview
  • source-preview
  • artifact-document
  • usage-meter
  • contextual-workbench
  • effective-policy-preview
  • memory-ledger-item
  • run-trace
  • handoff-packet
  • status-indicator
  • review-workflow as an evidence-review registry:block
  • approval-workflow as a human-in-the-loop registry:block
  • clarifying-workflow as a missing-decision registry:block
  • source-backed-artifact as a provenance-focused registry:block
  • memory-review as a durable-context registry:block
  • run-monitor as a background-work registry:block
  • multi-agent-handoff as an ownership-transfer registry:block
  • agent-settings as a policy-settings registry:block
  • workflow-phases as a phase-rail registry:ui
  • workflow-run-monitor as a multi-phase fleet-monitor registry:block

Validate the source registry locally with:

pnpm dlx shadcn@latest registry validate ./registry.json

Inspect the generated local JSON registry with:

pnpm dlx shadcn@latest list http://localhost:3000/r/registry.json
pnpm dlx shadcn@latest view http://localhost:3000/r/review-workflow.json
pnpm dlx shadcn@latest view http://localhost:3000/r/approval-workflow.json
pnpm dlx shadcn@latest view http://localhost:3000/r/clarifying-workflow.json
pnpm dlx shadcn@latest view http://localhost:3000/r/memory-review.json
pnpm dlx shadcn@latest view http://localhost:3000/r/run-trace.json
pnpm dlx shadcn@latest view http://localhost:3000/r/handoff-packet.json
pnpm dlx shadcn@latest view http://localhost:3000/r/run-monitor.json
pnpm dlx shadcn@latest view http://localhost:3000/r/multi-agent-handoff.json
pnpm dlx shadcn@latest view http://localhost:3000/r/agent-settings.json

Use the GitHub source-registry address for real installs. The generated local JSON is useful for previewing payloads while developing the site.

Quality Checks

Run the full verification suite before publishing changes:

pnpm run verify

CI runs this command on every PR and push to main. For visual QA, check both desktop and mobile breakpoints on the routes touched by the change. The current mobile bar is especially strict for composer, clarifying questions, citations, memory ledger rows, and template maps.

Design Direction

Agentic Craft is not a chat UI kit. It is a product-interface reference for agentic systems that need visible work, clear consequences, user control, recoverability, provenance, and reusable implementation patterns.

The design system favors:

  • compact mobile surfaces over oversized AI-demo chrome
  • source previews and citations close to the claim they support
  • locked consequence previews before approval
  • observable work instead of hidden reasoning traces
  • user-configurable prose and policy preferences
  • registry-ready primitives that can be composed into product workflows

About

No description, website, or topics provided.

Resources

Stars

Watchers

Forks

Packages

 
 
 

Contributors

Languages