Governments flooded by AI agents; CAPTCHAs are the worst response
As citizens batch-submit FOIA and subsidy applications through ChatGPT and Claude, governments instinctively add CAPTCHAs and block workflows, but the real answer is to redesign services so agents can handle the demand directly.
The video won't play here. Listen to the audio instead:
The argument · tap a timestamp to hear it
Citizens using AI move faster than governments can process applications
The host submitted only one FOIA request in seven years, then five in two weeks—purely because AI tools made it effortless. Lucas describes this as the core of ‘agentic flooding’: today, mostly people use LLMs to prepare materials before uploading, but soon agents will connect directly to government services natively. Most governments facing surging volumes respond by instinct: add CAPTCHAs and require offline declarations that LLMs weren't used. But that treats legitimate demand as an enemy.
— LucasGovernments need automation not to cut staff, but to replace retirees
Lucas points out the real constraint for OECD public sectors: not ‘overstaffed payrolls,’ but that vast populations of civil servants—doctors, teachers, back-office workers—average over 50 years old and face imminent mass retirement. The real question isn't whether automation will cause mass layoffs, but whether automation can keep pace with retirements and sustain baseline public services.
— LucasNew York's outdoor dining permits take 18 months but shouldn't
Take the New York City Comptroller's report on outdoor dining approval: nine approval steps, eight of which are pure form-filling and rework from bad 3D renderings—none require human judgment. Lucas's design: agents help applicants assemble and submit materials with pre-completion checks. Inside government, compliance checking runs on a fast track, completing in minutes, with edge cases reviewed by humans the same day. The only friction worth keeping is community notification—but even that can use VR walkthroughs and 3D renders for direct public participation, compressing 40 days of review to roughly 20.
— LucasPermits should be live database records, not four-year documents
Lucas proposes a counterintuitive redefinition: a permit is fundamentally a database entry recording legal authorization status and its conditions. If a condition becomes invalid mid-term—say, a safety problem emerges—don't wait four years for the next review; trigger a notification immediately. Conversely, if four years of data show all green, users shouldn't reapply; they just get a notification: ‘You've been operating successfully for four years.’
— LucasCitizens' shadow government services may become better than official ones
Lucas warns of one failure mode: when individuals or small teams using Claude Code and Codex exceed their government's capability, citizens spawn ‘shadow government’services that other nations copy—he mentions several countries have already built unofficial versions of USA.gov that everyone praises. This could either deepen discontent with official government or fill genuine gaps through good-faith citizens, depending on who builds it.
— LucasDigital government playbooks from the past decade won't work for AI agents
Lucas contrasts two eras. In the pre-agentic era, Estonia and Singapore spent 15–20 years perfecting a playbook—any new minister could copy it and know roughly what to expect in two or three years (Ukraine's Mykhailo Fedorov delivered results in two years; Greece's conservative government climbed from Europe's bottom quartile to top quartile in three or four). But the agentic AI era has no playbook yet. No one can precisely forecast ROI or user satisfaction. This is work for ‘natural experiments’in Mississippi, Estonia, and Ukraine to uncover.
— LucasWhen procurement experts code, even toilet paper gets better specifications
Concretely: several seasoned Estonian procurement professionals and one engineer spent months vibe-coding a small-purchase platform. It doesn't just write your procurement documents; it does market research—for toilet paper, the system flags ‘rolls’as vague and specifies that you should count by ‘linear feet’per 100 employees per year—detail nonprofessional buyers wouldn't know. Next, vendors can plug in via MCP server to participate in bidding; by year-end, small purchases should be entirely automated end-to-end. The key isn't the technology itself but that domain experts, not technologists, led product design.
— LucasOne governance tool solved what blocked state AI agents for months
Several state governments had built agents but held them back for months due to governance processes and uncertainty. After running them through a governance review tool that just launched with the Center for Civic Futures, problems that had stalled them for months dissolved in an hour. This shows grassroots vibe-coding is necessary but insufficient; the real bottleneck is missing organizational processes and governance frameworks.
— LucasIn their own words · checked verbatim
there is a very counterproductive likely reaction that I think a lot of governments will consider, which is, oh gosh, all of these inbound agents are terrible. If we're being flooded, let's raise the levees and let's damn the input shut.
Lucas4:58
We have decades of data you should be able to train on about like, you know, what makes, you know, the likelihood of someone overstaying their visa
we're very much pilled on the like, what's the nature of the AI going to be? And we're not talking so much about who's using the AI.
Lucas25:43
the big shit in this is not in any government specific is that code is no longer kind of the expensive bottleneck, right? But that just moves the bottleneck around.
Lucas35:40
what governments do, in fact, should be well specified in a public way. You know, a company has good commercial reasons to not open source everything they're doing... Whereas that's it's I mean, sort of a sauce of how government works. But it's like not only not shouldn't be secret, it should be public.
Lucas37:59
if the hard takeoff comes tomorrow, we're too late. But if in fact the hard takeoff or the takeoff is incredibly hard, it's a little bit more gradual, then who's building and the extent to which we can encode the sort of the societal level good outcomes that we care about into the algorithms is going to matter.
Lucas53:28
Glossary
- agentic flooding
- Citizens bulk-submitting applications and FOIA requests via AI agents, overwhelming government systems.
- MCP server
- A protocol that lets AI agents connect directly to external systems like government databases to perform operations.
How to listen
Entrepreneurs focused on GovTech and digital government, policy researchers, and anyone interested in how AI might reshape bureaucracy.
Skip transcription errors around 25:13 and the reggae theme at the end (after 58:43); low information value.