Playing with Claude Code (or its peers) is a revelation. Give them an idea, some context, and 20 minutes later you have a website, a mobile phone app, or a marriage-saving SaaS platform for determining whose turn it is to empty the dishwasher.
The way these tools work is by harnessing the raw power of a foundational model and turning it toward a specific purpose. The model supplies enormous general capability. The harness supplies context and direction, checks the work, and turns that capability into useful software.
This is already how software is being built at the technological frontier. Anthropic says Claude now authors more than 80 percent of the code merged into its codebase. At Google, AI generates nearly half. The humans have not disappeared. Their work has moved up a level: deciding what problems are worth solving, supplying context, directing the agents, reviewing their output, and determining what is safe and ready to ship. They operate the harness.
Government needs a harness, too.
Government does not lack people capable of writing code. It spends billions of dollars hiring companies that employ them. What it often lacks is the internal technical capacity to turn all that purchased engineering into working public services.
You do not need to recreate a giant software company inside every agency. You need a relatively small number of excellent technologists, positioned at the points of maximum leverage and given the authority to make decisions. They define the problem, translate policy into product direction, structure the work, manage technical tradeoffs, test whether the result works for users, and hold vendors accountable for outcomes rather than deliverables.
This is what makes New York City’s new PIT Crew so promising. The city is creating five small, multidisciplinary teams of product managers, designers, engineers, researchers, and data experts who will work inside government alongside city agencies. They are not buying another solution. They are building leverage: the institutional capacity to turn a mayor’s priorities, agency expertise, contractor labor, and modern technology into services that work for New Yorkers.
Colorado is following a similar path. The state is investing in technology with a starting place of the user’s experience, which allows for development that delivers outcomes, instead of just delivering technology systems. To do that, they spent years bringing technologists into government. Those designers, product managers, and engineers give Colorado the internal capacity to prioritize work. The state invested in expertise, not contracts.
We have seen the same dynamic at Federal Student Aid. When the FAFSA was on fire, FSA did not replace its large contractor workforce with thousands of federal engineers. It hired a few dozen experienced product, engineering, design, and delivery leaders and put them in positions of authority. That relatively small team helped redirect hundreds of millions of dollars in existing engineering capacity. The FAFSA stabilized, began shipping improvements continuously, and is now setting records.
Having this capacity and technical expertise works for more than just crisis response. When FSA discovered that over $1 billion a year of federal student aid was being stolen by organized criminals impersonating American citizens, Secretary McMahon directed FSA to verify the identities of anyone completing the FAFSA. While agencies in similar situations kicked off eighteen-month roadmaps to integrate new fraud detection solutions, FSA had live identity verification in place in three months. It was able to move this fast because product decisions were made by FSA employees who could reprioritize entire development teams without taking any contracting action.
The Trump Administration saw the leverage created by that relatively small pool of experts. It is now expanding the model across the broader student-loan portfolio, with as many as 200 additional technologists working inside FSA. Overall cost to the government has decreased. The Agency was already paying for this labor, it was just doing so through contracts with unaccountable vendors.
Treasury is doing the same. Under the leadership of Secretary Bessent, they are on an ambitious path to significantly reduce the number of outsourced engineers, buttressing new government employees with AI.
This is government’s new software harness: a small but authoritative layer of technologists inside public institutions, supplying the judgment, context, oversight, and relentless focus on outcomes needed to direct much larger teams of outside labor.
AI will only increase their leverage. The internal team that directs hundreds of contractor engineers today may soon direct contractors working alongside hundreds of software agents. The cost of producing code will continue to fall. The importance of knowing what to build, how its pieces fit together, whether it works, and who is accountable for it will rise.
Government can rent engineering capacity. It cannot outsource technical judgment.




I like this: "You do not need to recreate a giant software company inside every agency. You need a relatively small number of excellent technologists, positioned at the points of maximum leverage and given the authority to make decisions," and I am a fan of the ReCoding America agenda.
But also, think that it is relevant to keep the erosion of human agency in a highly automated, algorithmic decision-making regime in view at the same time. The government can have greater authority over tech, and therefore greater auditability and accountability, but it should focus on decision-making processes, not only outcomes. Could recoded America, now in expert command of its own technology, also turn to calibrating the use of algorithmic rule-making with other forms of deliberation?
By way of example, in an article I just finished (cite below), the example of the Arkansas Dept. of Human Services automated decision-making about disabled residents' care needs (which was previously done through a nurse visit). The algorithm determined that one foot amputee should have less home care because he did not have "foot problems." Rehousing technology should give agency heads more control not only over tech but also over the entire system and the role of automation within it.
(Amputee example came from: Ryan Calo & Danielle K. Citron, The Automated Administrative State: A Crisis of Legitimacy, 70 Emory L. J.
Available at: https://scholarlycommons.law.emory.edu/elj/vol70/iss4/1)
Asked Opus to scrape data and to write Rmd to create a database and some tibbles and figures from the database.
Wrote text in Rmd using the database and figures generated by Opus.
Send code and the rendered files to DeepSeek to take them apart.
DS finds a few big mistakes, not hard to fix, and suggests hundreds of small changes, which I work thru one by one.
This is not advanced stuff I’m doing.