Stack mapping

Every item on the hackathon kick-off slide, where it lives in this repo, and where a judge can see it during the 3-minute demo. If it is not visible on screen, it does not count.

Kick-off item	Where it lives	Visible in the demo as
MCP	MCP tool server in mcp_server/ (8085); called by Finder, CDR and Pipeline	"MCP tool calls" panel, bottom left — search, places, rag_retrieve, persist_and_schedule, inbox with their arguments
AG-UI	ui_client/agui/ (CopilotKit) + POST /ag-ui on 8084	The artifact drawer. Agent tool calls mount real components — research brief, content package, critique card, email, calendar — not chat text
OpenTelemetry	Spans named {agent}.{pattern} on every agent	The live agent trace. Each row is a span: agent name, pattern badge, service, summary
LangGraph	P2's graphs in the CDR service	Sequential and loop badges — ResearchThenPropose, and the RefinementLoop that fails a draft then passes the rewrite
DeepAgents	CDRRootAgent, the root planner — DeepAgents-style (plan → delegate to subgraphs as tools), hand-written on LangGraph, not the deepagents package	First line of every run: "DeepAgents root … delegating to finder, research, package, outreach subgraphs"
Claude Agent SDK	Optional specialist critic in harness/claude_agent.py	FactCheckerAgent, when ANTHROPIC_API_KEY is set. GET :8084/health reports claude_agent_sdk.available so a dead integration is visible before a demo, not during one
AWS Bedrock AgentCore	Deploy target for the recorded demo	Not on screen — deployment target, called out verbally
Groq	Local inference for the llm-pattern agents	Green llm badges: DraftWriterAgent, ProposalGenerationAgent, PitchEmailAgent, ReplyClassifierAgent

Agent patterns on screen

The trace badges the pattern for every agent, because "which pattern is this" is the question judges actually ask.

Badge	Meaning	Count	Example
llm	Single model call	26	DraftWriterAgent, FactCheckerAgent, PitchEmailAgent
sequential	Ordered pipeline	5	ResearchThenPropose, OutreachPipeline, PersistAndSchedule
custom	Hand-written control flow	2	CDRRootAgent, OpportunityFinderRoot
parallel	Fan-out / gather	1	ParallelResearch — four research specialists at once
loop	Iterative refinement, max 3	1	RefinementLoop — fact + voice critique, rewrite, re-check

35 distinct named agents are defined across the four services and appear on the trace by name. Verify the count yourself:

grep -rho 'name = "[A-Za-z]*"' */agents/*.py | sort -u | wc -l

Most carry the llm badge because most are a single prompted step. The patterns that matter for the brief — parallel fan-out, and a refinement loop that visibly fails then passes — are ParallelResearch and RefinementLoop, both of which run in every campaign.

Human-in-the-loop

Deliberately minimal. One Run campaign click is the only required human action in the entire demo. PAUSE_BEFORE_SEND exists as a toggle and defaults off — outreach sends without approval. There is a Stop button; it ends the run, it does not gate it.

Protocol note

The UI never learns whether events came from a fixture file or from P2's live agent — both arrive as the same AG-UI SSE stream from POST /ag-ui. Going live is USE_FIXTURES=0, not a rewrite. See demo/fixtures/README.md for the event contract and cdr/agui_map.py for the one module that holds P2's schemas to it.

If the live CDR stops answering mid-run, the board says so on the trace and keeps the events it already drew; if it never answered at all, the run falls back to the fixture replay and labels itself. Either way the demo does not hang.