AI Agents

Agent workflow samples, latest-result receipts, and dispatch controls. Inspect the outputs first; queueing a fresh run depends on the configured bridge or GitHub workflow permissions.

Home Start Here Products Chat with Polly Agents GitHub

New here? Read this first.

An agent is a bounded workflow that does one job and reports back with a receipt. The public page can always show the latest published results. Fresh dispatch requires the Vercel bridge or GitHub Actions path to be configured.

The simplest use: click a starter card below to load a sample task, then queue it if the dispatch bridge is available. If it is not, the page opens the GitHub workflow fallback instead of pretending the run happened.

If you'd rather have help picking the right tool for a bigger task, open Polly chat or see the three starter routes.

Load a sample workflow

Each card pre-fills the dispatch form below and scrolls you to the queue button. These are workflow recipes, not proof that a fresh run has already completed.

📚 Research a topic

Asks the research agent to find recent sources on a topic, read them, and produce a structured report with citations.

Pre-fills with: "Latest news about AI safety governance"

💬 Ask a question

Web-searches for context and produces a plain-language answer with source links. Good for "explain this to me" questions.

Pre-fills with a sample question

🔭 Watch a site

Quick-reads one or more URLs and reports title, word count, link count, and a content preview. Useful as a smoke test.

Pre-fills with two example URLs (comma-separated)

🪟 Read a page

Pulls structured data from any URL — text, links, tables, headings, metadata. No AI involved, just a clean read.

Pre-fills with a Wikipedia URL

Developer? The full agent catalog and recipes are below. Skip the friendly cards.

Available Agents

Research Agent

Searches the web, reads multiple sources, scores relevance, and produces a structured report with confidence rankings when the configured runner is available.
Receipt-backed Configured runner Web search

Site Monitor

Quick-reads multiple websites and reports title, word count, links, and content preview. Runs every 6 hours automatically.
Free Playwright Scheduled

Ask Agent

Answers a question using web context and the configured model path, then publishes the answer with source citations and run metadata.
Receipt-backed Model path varies Ollama/HF capable

Page Analyzer

Scrapes any URL and extracts structured data: text, links, tables, headings, metadata, JSON-LD. No AI needed.
Free Playwright

Coding Smoke

Runs the SCBE coding-system benchmark lane and returns pass/fail evidence for agentic coding readiness.
Local Benchmark

System Build Smoke

Checks router entrypoints, required scripts, Node syntax, and Python compile health before heavier build promotion.
Fast CI-ready

Agentic Benchmark Ladder

Runs Level 0 smokes plus Level 1 repo-native tasks; optional Level 6 adds GeoSeal/CLI surface pytest gates. Reports multi-metric rollup (time, secret-leak heuristics, pass rate) instead of a single score.
Ladder Metrics

Pair Benchmark

Compares a one-body coding workflow against the paired builder/navigator pattern on SCBE-native coding tasks.
Zero-cost Dual-agent

Poly Coding Seed

Regenerates the external coding seed corpus with language lenses, binary/hex transport, CA operation hints, and GeoSeal metadata.
Training GeoSeal

Agent Bus

Dispatches one SCBE agent-bus event through scripts/system/agentbus_pipe.mjs. Bridge between the published scbe-agent-bus npm/PyPI package and the website chat surface; returns a typed envelope.
Local scbe-agent-bus Typed envelope

Run a Task

Quality Harness

What is measured now

The page reads the same JSON the agents publish: task freshness, checked sources, accepted findings, confidence spread, source outcomes, errors, and output hashes. This makes weak runs visible instead of hiding them behind a clean-looking summary.

Known limits

GitHub Actions can be delayed, scraped sites can block bots, model providers can fail, and the current public page shows latest results only. Historical run comparison needs a retained run archive before it can be shown honestly.

Latest Results

Loading agent data...

How It Works

  1. You submit a task — the button tries the configured Vercel bridge first
  2. Fallback path opens — if the bridge is unavailable, the page opens the GitHub Actions workflow
  3. The runner gathers context — search, scrape, benchmark, or local bus logic runs according to the selected task
  4. The configured model path processes context — this may be HuggingFace, Ollama, or a non-LLM scraper depending on the task and credentials
  5. Results publish here — the public page reads the latest JSON receipt once a run has actually produced one

The heavy agent run still happens on GitHub Actions. A small Vercel bridge can queue runs and read status when configured; otherwise this page falls back to the manual GitHub workflow. The SCBE governance pipeline gates every action through the 14-layer security pipeline.

Hire Issac