Closed Is a Stage, Not an Ending: How AI Agents Keep Our Funding Dashboards Honest

In mid-July, Missouri’s Department of Economic Development opened a public comment period on its Opportunity Zone nominations. There was no press release. The state’s OZ landing page looked unchanged. The announcement went out as a GovDelivery email bulletin, with the documents parked at media-file URLs and the survey hosted on Qualtrics. If you were watching the official page — as our automated monitor was — you saw nothing.

We caught it late. That miss, and a few like it, reshaped how every one of our dashboards works.

Key takeaways

• Prosody Labs runs six public trackers — federal discretionary funding, Opportunity Zone 2.0 activity across the states, the SHIPS for America Act, USDA’s FIELDS program, and award analytics for PIDP and the Marine Highway Program — with no one refreshing them by hand.
• Scheduled Claude agents research each beat and rewrite structured data files; an n8n workflow commits the results to GitHub and the site redeploys itself.
• The engineering was the easy part. The editorial doctrine — what to watch, where news breaks, when to distrust an unchanged webpage — took real misses to learn.
• The core rule is now written into every agent: closed is a stage, not an ending. A passed deadline is not a reason to stop watching.

What the dashboards are

Six trackers, public at prosodylabs.prosodyconsulting.com. Fast-moving programs get daily runs; slower ones get weekly. Presentation and data are strictly separated: each agent touches only a structured JSON file of facts, never the page around it. When the agent finishes, an n8n workflow commits the new data to GitHub and Cloudflare rebuilds the site. No analyst opens a spreadsheet.

Describing the pipeline takes one paragraph. Making it trustworthy took a doctrine.

The rules the agents had to learn

Early on, our Opportunity Zone monitor missed three state developments in a row — a Treasury filing in Arizona, a draft-list comment window in Vermont, an extension request in Maryland. All three happened after the tracked deadlines had “passed,” so the monitor had stopped looking. Reasonable behavior for software. Wrong behavior for an analyst.

Closed is a stage, not an ending. An item now leaves the watch rotation only at a true terminal milestone — zones designated, awards obligated, a bill enacted or dead at the end of a Congress. Everything between close and terminal gets tracked explicitly: draft lists, comment windows, extension requests, early publications.

Official pages confirm; they don’t break. Agencies announce through GovDelivery blasts, governor press offices, practitioner newsletters, and trade press — often before their own websites change, sometimes instead. Nebraska’s Treasury filing in July surfaced only on the governor’s press page; the agency site said nothing for days. Every run, the agents sweep those channels, not just the official URLs.

Silence must be distinguishable from neglect. Any row not re-verified within its window gets an explicit re-check, and an unchanged row gets stamped “re-verified, no change.” A quiet dashboard should mean we looked and nothing moved — not that nobody looked.

Credible secondhand intel goes in immediately, with attribution and an “awaiting official posting” flag. It gets upgraded when confirmed and removed if it isn’t. Waiting for the official page means being last.

None of this came from a model. It came from writing down how a good analyst actually works and holding software to it, every run, without days off.

What it means for you

• If you’re tracking a program where the deadline has “passed,” don’t stop. Comment windows, draft lists, and appeals routinely open after close — and they’re often the last chance to influence an outcome.
• Subscribe to the email bulletins (GovDelivery and similar) for every agency you care about. That’s frequently where news breaks first, and sometimes the only place it breaks at all.
• If you’re building AI agents for monitoring work, budget more time for the editorial rules than for the plumbing. The plumbing is a solved problem; judgment is not.

Related reading

• 35 jurisdictions, 64+ consecutive runs, not one missed day
• Prosody Labs — the live trackers
• Technology practice

How we read it: the value of AI in this kind of work is not speed — it’s coverage that never lapses, governed by rules a senior analyst would recognize. If you want to know whether an autonomous tracker fits your problem, tell us about it and within a week we’ll say whether it’s a fit, what we’d build, and what it costs to run.

This is policy and market intelligence, not legal advice. Talk to us about your project →