How the Government Could Use Prize Challenges to Spur True Agentic AI Innovation

Prize-based acquisition to stimulate innovation isn't new. Federal agencies have had broad prize competition authority since the America COMPETES Reauthorization Act of 2010.^[1] MITRE has argued for years that challenge-based acquisition can outperform the traditional single-performer RFP model in some settings.^[2] NASA's Center of Excellence for Collaborative Innovation has run crowdsourced problem-solving for over a decade.^[3] And DARPA's AI Cyber Challenge just proved, at real scale, that a government-run AI prize competition can produce working autonomous software.^[4] This post is about a variable those programs weren't designed to test: the price. What I want to propose is setting the prize deliberately below what a conventional bid would cost — and using that gap to find out what agentic AI teams can actually do.

The Pilot Project

Here is the setup. The government establishes a pilot project. The project contains a set of open source, unclassified software needs framed as prize challenges. The needs are defined by government analysts and software engineers. At a minimum, each challenge would contain:

• An open source data set of interest (at least initially),
• A set of use cases and requirements for software that uses the data,
• A cloud environment developers may work in,
• A limited budget via OpenRouter for any model developers wish to use,
• An API and interfaces for the developed software,
• A set of unit and functional tests,
• A target deployment environment (the hardware, processor, etc.) where the finished product will live,
• Target performance metrics,
• Security tests and harnesses, and
• A hidden evaluation set to test against.

Here's the last piece: the prize, set deliberately low. If the honest estimate is $250K even with agentic AI, the prize is $150K — 40% under.^† It's an intentionally aggressive starting point that establishes a reward meaningfully below the conventional cost estimate — low enough that winning likely requires a different production method rather than merely a cheaper conventional bid.

That's not the standard "prizes are cheaper than contracts" argument. This is a narrower, specific bet: that agentic AI has already cut the cost of building working software by more than the market has priced in. The fastest way to find out by how much is to dare a team to prove it under a number that makes conventional estimators uncomfortable. Note, the pilot projects really need to have those attributes listed above well-defined and clear!

This sort of looks like a firm-fixed-price (FFP) contract, but it isn't. Under FAR 16.202-1, an FFP contract establishes a price that is not adjusted based on the contractor's cost experience, placing responsibility for cost, profit, and loss on the contractor.^[5] Here, teams do the work before they know they'll be paid — the government pays for demonstrated success, not a promise to perform.

Another benefit: the government doesn't need to predict which kind of organization can build the software — a prime contractor, a startup, a university lab, or three engineers who've never worked together but know how to orchestrate agentic AI well. Traditional development carries the cost structure of whoever delivers it: management, overhead, proposal teams. A prize challenge lets a completely different production model compete on the same terms, and the tests — not the pedigree — decide who wins.

A Cheaper Experiment

AIxCC is probably the closest precedent to what is proposed above. DARPA spent two years and offered a cumulative $29.5M in prizes asking whether AI could autonomously find and patch vulnerabilities — a genuinely open capability question, funded generously because the answer was unknown.^[4]

This pilot points the other way. It doesn't fund a team generously to push the frontier; it sets an intentionally aggressive prize against any type of problem, from researchy to the mundane. It just needs a well-understood requirement that enables people to see what they can do with AI that already exists. This tests the labor model — and it's a cheap experiment to run, since it needs no multi-year program, just an agency willing to set an aggressive reward and see who shows up.

The prize authority already exists. The technology already exists. What's missing is an agency willing to set the number lower than conventional estimates suggest is reasonable, and find out who was right.