Add incumbent-on-post example: agent + MCP, scorer, evals

[makegov-mark]

What

A new example, examples/incumbent-on-post/: when a new opportunity hits SAM.gov, brief the likely incumbent, the prior award value, and whether the money is still flowing — from one notice.

It's the cookbook's first MCP integration. Every other example hand-wires the tango-python SDK; here the tool catalog comes from the hosted Tango MCP server (https://govcon.dev/mcp) and a PydanticAI agent drives it.

Why this shape

The obvious join — match the new opportunity to the prior contract on the solicitation number — doesn't work: a fresh recompete gets a brand-new number; the predecessor award carries the old one. There's no clean key. So the recipe pivots on the contracting office (the thing that persists across a recompete) and ranks candidates by whose award is expiring near the notice — not by who won most recently. Because that's an inference, the output is a ranked shortlist with confidence + evidence, never a single asserted name, with "likely new work" as a first-class answer.

The files

File	Role
brief.py	PydanticAI agent over the hosted Tango MCP → typed IncumbentBrief (the easy version).
scorer.py	Deterministic candidate ranking (expiry-window × dollars × recurrence), bounded server-side with expiring_gte/lte — reproducible, eval-able (the honest version).
evals.py	Structural invariants on the scorer; runs without an LLM key.
sample_notice.json	A realistic enriched notice to run against.
demo.svg	Self-contained pipeline diagram, built from a real run.

Plus: just incumbent / just incumbent-evals, pydantic-ai-slim[anthropic,mcp,openai] added to the examples dependency group, and both README tables updated.

Model: hosted or local

The agent uses the hosted Anthropic model by default, but runs against any OpenAI-compatible server (LM Studio, Ollama, vLLM) by setting OPENAI_BASE_URL + MODEL — at which point ANTHROPIC_API_KEY isn't needed. The MCP server (hosted vs. local stdio) is an independent axis. Both are documented in the example README.

Testing

• just incumbent-evals — all scorer invariants hold against the live API (top court-reporting incumbent at VA NCO 5 resolves to Vet Reporting LLC; far-future window correctly returns "new work").
• just incumbent — agent connects to the hosted MCP and returns a structured 3-candidate brief, independently agreeing with the scorer's top pick, and catching a real office-code disambiguation (36C250 vs 36C245).
• demo.svg validates as XML.

Per repo convention, this example is not run in CI (it calls a paid LLM and a live API); evals.py is deterministic but still hits the live API.

Risks / notes

• demo.svg hardcodes the sample brief's specifics (vendor names, 2029 recompete date, −3% funding figure). If sample_notice.json changes, the diagram drifts from a fresh run — kept real for impact; easy to genericize if preferred.
• The funding signal is wired through the schema/prompt but is the lightest part of the agent's job; deepening it with get_budget_account + quarterly cash flow is a documented next step.

🤖 Generated with Claude Code