P2 · ACT WITH TOOLS UNDER BUDGET

An agent that stops, and can tell you which rule stopped it.

LIVEv1.0.0 · main 1b4915f · CI greenVerified 2026-08-14

A plan → retrieve → critique loop where the budget lives in state rather than in the prompt, so termination is a property of the program instead of a request to the model. Every run ends in a typed status — done, refused, budget_exhausted, capability_missing, backend_unavailable — and emits a deterministic trace of six event types ending in stop. Hire this if you have been burned by an agent that ran until someone noticed.

HOSTED · no clone, no key

curl https://pax-agentic-rag.vercel.app/health

No clone, no key. Ask a question the fixture cannot support and read the typed refuse. Hosted health reports version 1.0.0; GET /health and GET /metrics both returned 200 on 2026-08-14.

LIVE

Every line here is backed by a test, a route, or a file in the repository at v1.0.0 · main 1b4915f · CI green.

  • Plan → retrieve → critique loop with a retrieval budget and a no-repeat bound held in ResearchState.
  • Typed terminal statuses, with refusal and dependency failure kept apart: a missing backend is backend_unavailable, never an evidence refusal.
  • A complete deterministic trace of six event types ending in stop, exportable from the API and from the result page.
  • POST /v1/research, a CLI, GET /health, and GET /metrics in Prometheus text — all four verified over HTTP on the hosted instance.
  • A 17-case golden evaluation across five slices, run with empty provider variables.
  • An optional production-rag HTTP retriever, opt-in behind PRODUCTION_RAG_URL and mock-transport tested.
  • Stored run artifacts with GET /v1/runs, SSE stream, and smoke-tested download offsets at main 1b4915f (fixture-only on the hosted copy).

NOT SHIPPED

The boundary, stated before you find it yourself. Nothing below is presented as built, hidden in a footnote, or described as coming soon.

  • The hosted demo runs the committed fixture retriever. It proves the agent's control flow and claims nothing about retrieval or answer quality.
  • No PRODUCTION_RAG_URL is set on the hosted instance, so the HTTP retriever there fails with a typed capability_missing by design rather than falling back.
  • No paired agent-uplift measurement against a single-pass baseline. That is the next quality boundary, and it has not been run.
  • No model-based planner, auth, rate limiting, streaming, multi-tenancy, load figure, or arbitrary web or write tool.
  • The optional HTTP adapter is mock-transport tested; this release claims no live cross-service measurement.

Capture

One image from the repository at the commit named in the caption.

Agentic Research UI showing a completed run with its stop reason and cited synthesis
A completed research run: budget spent, stop reason, cited synthesis. Source file

INTERVIEW THIS SYSTEM

Three questions worth asking, and the answers I would give.

Published in advance, because a trade-off I can only defend when nobody has read it first is not a trade-off I understand.

  1. 01

    Why does the step budget live in state instead of in the system prompt?

    A prompt-level budget is a request the model may decline. A state-level budget is a loop invariant the model cannot see or argue with, so termination is testable without invoking a model at all.

  2. 02

    Why is an unreachable backend not a refusal?

    Refusal is a claim about evidence: the corpus does not support an answer. An unreachable backend is a claim about infrastructure. Collapsing the two teaches an operator to ignore refusals, and hides the outage.

  3. 03

    What does a 17-case fixture evaluation actually prove?

    Control-flow conformance: that each slice reaches the terminal status the contract says it should. It is plumbing evidence. Anyone reading it as a retrieval quality number is reading a claim the harness explicitly refuses to make.

Keys

How to turn the pages

On
15
Open a chapter
g h
Home
g i
Interview kit
g s
Studio
/ or CtrlK
Jump: type a word, not a command
?
This page
Esc
Close it again

Chapters

  1. 1
    01 Production RAGRetrieve and answer honestly
  2. 2
    02 Agentic ResearchAct with tools under budget
  3. 3
    03 Multi-Agent OrchestrationCoordinate specialists
  4. 4
    04 RepoMindUnderstand codebases
  5. 5
    05 AI PlatformOperate the system

Keys sleep while a field has focus, so typing a g into the jump box types a g.