Most AI programs die between the pilot and the auditor. The work in between — sitting with
the people who do the job, writing down what actually happens, redesigning it, then building
the integrations into the systems you already run — used to take a consulting engagement.
With agent harnesses it takes a fraction of that. Palantir's forward-deployed engineers have
worked this way since 2005; Chamath Palihapitiya's 8090 raised $135M building a software
factory on it. The scarce part was never the harness. It is modeling the domain
correctly — which is twenty-five years of my working life.
Right now this is a pro bono practice. One nonprofit at a time, at no charge,
on evenings and weekends. That is arithmetic rather than charity: the cost of this work fell
far enough that one person can do it properly for free. No fees, no proposals, nothing to buy.
All of it unpaid, one organization at a time. I'd rather do one thing properly than three things badly.
An honest read on where AI would genuinely help you, where it would waste money you don't have, and what to do first. Including the answer nobody selling it will give you, which is sometimes “not yet — fix this instead.”
What staff and volunteers may and may not put into these tools, how donor and beneficiary data is handled, and what you tell a funder who asks. Written to be read by non-technical trustees, not by lawyers.
Someone in your corner for the architecture and the contract — what the system actually does, where your data goes, what happens when you want to leave, and whether the price makes any sense.
Platform choices, what to consolidate, what to retire, which cloud spend to stop. The unglamorous work that decides whether anything else you fund actually lands.
Palantir has embedded engineers this way since 2005. 8090 is building a business on it with $135M behind them. What changed recently is that one person can run the same loop for one organization. Here is what it does, and exactly where it stops.
Interviews, recorded working sessions, and the systems themselves. Tooling can now sit across speech, slides and diagrams and answer questions about them with the evidence timestamped — I co-authored a peer-reviewed system that does this for technical meeting video.
Where it stopsIt captures what people say happens. Noticing what they left out — or what nobody wants to say while their manager is in the room — is still a human job.
The description becomes an explicit model: the steps, the decisions, the handoffs, the data that moves between them. In Palantir's version this is ontology modeling — mapping a customer's systems into objects, links and actions — and it is the step their own engineers lead rather than hand over. It is also just enterprise architecture — the thing I have spent a career on, rather than a skill picked up when this became fashionable.
Where it stopsI co-authored the measurement of this one. A critique-refine loop genuinely does improve correctness and completeness over a single pass — but when the model checks its own structure, the results “often still show structural inconsistencies.” Something deterministic has to sit in the loop. That seam is precisely where a pilot passes and an audit fails.
An agent will faithfully optimize whatever process you describe to it. It will not tell you the process should not exist, that two teams have been duplicating the same work for three years, or that the form has fourteen fields because of a grant condition from 2011 that expired.
Where it stopsThis is judgment, politics and institutional memory. It is the step everyone skips, and skipping it is how organizations end up with a faster version of the wrong process.
Agents reach real systems through tool interfaces, with explicit permission gates, deterministic checks that run before anything executes, read-only enforcement where it matters, and hard budget ceilings. Every action it takes is logged.
Where it stopsAll of that is opt-in. Pointed at your donor database without those gates, an agent will do damage faster than a person could. The same harness that makes the work quick is what makes it auditable — but only if someone configures it that way, and that is the whole job.
Both 8090 and Palantir will also sell you the platform you end up standing on. Palihapitiya's own public warning is that AI adopted without a coherent strategy recreates exactly the vendor problem SaaS already gave you. I have nothing to sell you, which is the only reason my read on any of it is worth having.
Open source, peer-reviewed research, and a granted patent — built and maintained personally, on my own time. Two of these are the agent-orchestration layer the method above runs on; two of the papers measure where it breaks. This is the part of the record that is mine to show you.
Donor records, case management, beneficiary information. You are being told AI will help, and you need someone who has run governance in a regulated environment to tell you what is safe to try and what is not.
You are the person who has to decide, and there is nobody in the building to ask. Sometimes you need thirty minutes with someone who has made the same decision before and has nothing to sell you.
One or two people holding up everything, with a platform decision in front of them they cannot easily reverse. A second opinion from outside, with no stake in the answer.
I have chaired the Association of Nepalis in the Americas' annual national convention since 2009 and advised A New America Foundation. I know what volunteer capacity actually feels like, and I will not hand you a plan that assumes a team you do not have.
Tell me what's actually broken. There is nothing to buy here — if LapaxWorks isn't the right answer, I'll say so and point you somewhere better.