production-ready-agent

Deal Desk — a Claude Managed Agents workshop

Workshop sample code. Not maintained and not accepting contributions.

A chat-first UI over a multi-agent M&A research team, built on the Claude Managed Agents API. A coordinator agent delegates to four research sub-agents in parallel, reads prior-deal lessons from a memory store, can reach Linear via MCP + Vault, and emits a graded investment thesis. The UI streams every event, lets you confirm gated tool calls, define outcomes inline, and tab into each sub-agent's thread.

All companies and financials are fictitious.

Repo layout

starter/      The app you build during the workshop. The UI, event reducer,
              and provisioning are done — the API route handlers are stubbed
              with TODOs. Start here.
solution/     The finished reference. Diff against starter/ when stuck.
seed/         Shared workshop assets: agent configs, sandbox env, memory
              entries, fictional target CSVs, the outcome rubric.
bin/          Idempotent provisioning scripts that hit the Managed Agents API
              once and write IDs into .env. Both apps read the same .env.

starter/ and solution/ are identical Next.js apps; the only difference is whether the route handlers in app/api/ are filled in. Diff them any time:

diff -ru starter/app/api solution/app/api

Prerequisites

• An Anthropic API key with the Managed Agents beta enabled
• The ant CLI: brew install anthropics/tap/ant on macOS, or see the quickstart for Linux and other platforms
• bun and jq

Setup

Provision once — both apps share the resources:

cp .env.example .env
# Fill in ANTHROPIC_API_KEY.

./bin/setup.sh               # provisions environment, sub-agents, coordinator
                             # (v1, no roster), memory store, files; writes
                             # the IDs back into .env and copies it into each app
./bin/enable-multiagent.sh   # adds the sub-agent roster to the coordinator → v2
                             # (you'll run this mid-workshop, but it's safe now)

cd starter && bun install

setup.sh is idempotent — re-running reuses existing IDs instead of creating duplicates. The coordinator is the exception: it is recreated fresh at v1 each time, because agents.update cannot remove an existing multiagent block. Sessions always use the latest agent version, so once enable-multiagent.sh has run, new sessions get the sub-agent roster.

Run

cd starter && bun dev
# open http://localhost:3000

Each app reads its own .env (synced from the repo-root .env by setup.sh). lib/anthropic.ts reads the file directly so a shell-exported ANTHROPIC_API_KEY won't shadow it.

What you'll build

starter/README.md walks the seven exercises. In short:

TODO	File	API surface
1. List sessions	app/api/sessions/route.ts	sessions.list
2. Create a session	app/api/sessions/route.ts	sessions.create — agent, environment, resource mounts, vault
3. Send messages and outcomes	app/api/steer/[id]/route.ts	sessions.events.send — user.message, user.define_outcome
4. Retrieve a session	app/api/session/[id]/route.ts	sessions.retrieve
5. Bridge the event stream	app/api/stream/[id]/route.ts	sessions.events.list + sessions.events.stream → SSE
6. Confirm gated tool calls	app/api/confirm/[id]/route.ts	sessions.events.send — user.tool_confirmation
7. Delete a session	app/api/session/[id]/route.ts	sessions.delete

Already done for you and worth reading:

Feature	Where
Per-thread event streams	app/api/stream/[id]/thread/[threadId]/route.ts
List sub-agent threads (threads.list)	app/api/session/[id]/threads/route.ts
SSE encoder / dedup / error bridge	lib/sse.ts
Event → chat message reducer	lib/chat.ts
Agent + sub-agent definitions	seed/agents/*.yaml
Sandbox config	seed/environments/research.yaml
UI	components/ChatPanel.tsx, SessionRail.tsx, ai/*

Optional: MCP + Vault via Linear

The coordinator can read open diligence issues from Linear over MCP, with the end-user's Linear token held in a Vault. To exercise it, fill LINEAR_API_KEY and LINEAR_TEAM_ID in .env, run ./bin/seed-linear.sh to create some fictional diligence issues, and create a Vault holding the same token. Set VAULT_ID and toggle "MCP" on in the New Session modal.