I work in a very similar field and found this fascinating. We’re looking to keep our development team fairly traditional, as in writing code by hand but using AI to empower our non tech staff with the tools to build their own interfaces.
Honestly, I think you’re leaving a huge amount on the table by doing it that way.
The productivity I have right now, even in this semi-automated state, is ridiculous. It genuinely feels like I’ve accidentally acquired the powers of a minor software god.
I can work across four software projects at once, any one of which would probably have needed three or four people not very long ago. Now I can do it largely on my own, with one AI acting as architect, another as implementer, another as tester, and another standing there with a clipboard saying, “Yes, but this doesn’t actually test the thing you think it tests.”
That last bit is the killer.
The antagonistic-agent model is where it gets really interesting. You don’t just get code faster. You get code constantly challenged. “This test misses that scenario.” “This edge case isn’t covered.” “This implementation is too tightly coupled.” “This looks clever but will be horrible to maintain.”
The result is that everything trends towards very high coverage, but more importantly, useful coverage. Not decorative coverage. Not “look, the number went up” coverage. Actual adversarial testing of the thing you’re building.
And once I get the weighted trust model working properly, so I’m no longer the human bottleneck in every decision, I think it’s going to get genuinely mind-blowing.
I think I shall have a chat with you about it at some point, I’ve barely scratched the surface because I’m always being chased by people. Just need a few days and could set up a similar toolkit and bamboozled people.
I like the overall design of Relay. Reminds me of the Engineering Department and saner bits of Product Marketing back at UstaCorp. (I retired in 2008, long before AI, when I felt I might be aging out of the sooperdooper software staffer role.) I find it interesting that AI agents can take on project management roles over other task specific agents, but it is obvious in hindsight.
The whole architecture of orchestrating agents and task agents linked with standardized protocols and language feels almost biological, reminiscent of how the subunits of a human brain work together.
At first glance, Relay looks like a bit of workflow plumbing: pass this instruction here, send that result there, have one agent review another agent’s work, push something into VS Code, pull something back out again.
But the deeper pattern is much more interesting than that.
The moment you stop thinking of AI as “one clever box I ask questions of” and start thinking of it as a set of specialised cognitive functions, the architecture changes completely.
One agent can hold the strategic shape of the work. Another can implement. Another can test. Another can criticise. Another can watch for drift. Another can maintain context. Another can say, “Yes, very clever, but that is not actually what we agreed to build.”
That does start to feel oddly biological. Not because I think it is literally a brain, obviously, but because the pattern becomes less like a single expert and more like a nervous system: signals, inhibition, correction, reinforcement, escalation, trust, memory, attention.
And that, to me, is where this gets properly interesting.
The magic is not just in making one agent smarter. It is in making the whole system less stupid.
That is also why I’m so focused on protocols, contracts, hand-offs and trust models. Without those, you just have a room full of very enthusiastic interns shouting at each other. With them, you start to get something closer to coordinated work.
One thing that does make me laugh, though, is that even in my synthetic organisation, I have somehow recreated the class politics of enterprise software.
All the senior synthetic people get proper names.
Henry is the architect. Margo is the red team lead. They have personalities. Roles. Gravitas. Opinions. Probably opinions about wine.
And then, down in the implementation mines, all the poor software engineers are just called Mini-Claude.
Which Mini-Claude?
Who cares?
They’ve got GUIDs against their names. That’s plenty. Get back to work.
So yes, Relay may be synthetic, but apparently I have still managed to reproduce the sacred corporate hierarchy: the architects get names, mythology and strategic intent. The engineers get tickets, identifiers and a vague sense that somebody upstairs has had a “vision”.
Still slightly mad, of course.
But then again, most useful software architecture is just madness with interfaces.
I work in a very similar field and found this fascinating. We’re looking to keep our development team fairly traditional, as in writing code by hand but using AI to empower our non tech staff with the tools to build their own interfaces.
Honestly, I think you’re leaving a huge amount on the table by doing it that way.
The productivity I have right now, even in this semi-automated state, is ridiculous. It genuinely feels like I’ve accidentally acquired the powers of a minor software god.
I can work across four software projects at once, any one of which would probably have needed three or four people not very long ago. Now I can do it largely on my own, with one AI acting as architect, another as implementer, another as tester, and another standing there with a clipboard saying, “Yes, but this doesn’t actually test the thing you think it tests.”
That last bit is the killer.
The antagonistic-agent model is where it gets really interesting. You don’t just get code faster. You get code constantly challenged. “This test misses that scenario.” “This edge case isn’t covered.” “This implementation is too tightly coupled.” “This looks clever but will be horrible to maintain.”
The result is that everything trends towards very high coverage, but more importantly, useful coverage. Not decorative coverage. Not “look, the number went up” coverage. Actual adversarial testing of the thing you’re building.
And once I get the weighted trust model working properly, so I’m no longer the human bottleneck in every decision, I think it’s going to get genuinely mind-blowing.
Honestly, you should speak to me about it.
I’m getting pretty slick at this now.
I think I shall have a chat with you about it at some point, I’ve barely scratched the surface because I’m always being chased by people. Just need a few days and could set up a similar toolkit and bamboozled people.
I like the overall design of Relay. Reminds me of the Engineering Department and saner bits of Product Marketing back at UstaCorp. (I retired in 2008, long before AI, when I felt I might be aging out of the sooperdooper software staffer role.) I find it interesting that AI agents can take on project management roles over other task specific agents, but it is obvious in hindsight.
The whole architecture of orchestrating agents and task agents linked with standardized protocols and language feels almost biological, reminiscent of how the subunits of a human brain work together.
That is exactly the bit that fascinates me.
At first glance, Relay looks like a bit of workflow plumbing: pass this instruction here, send that result there, have one agent review another agent’s work, push something into VS Code, pull something back out again.
But the deeper pattern is much more interesting than that.
The moment you stop thinking of AI as “one clever box I ask questions of” and start thinking of it as a set of specialised cognitive functions, the architecture changes completely.
One agent can hold the strategic shape of the work. Another can implement. Another can test. Another can criticise. Another can watch for drift. Another can maintain context. Another can say, “Yes, very clever, but that is not actually what we agreed to build.”
That does start to feel oddly biological. Not because I think it is literally a brain, obviously, but because the pattern becomes less like a single expert and more like a nervous system: signals, inhibition, correction, reinforcement, escalation, trust, memory, attention.
And that, to me, is where this gets properly interesting.
The magic is not just in making one agent smarter. It is in making the whole system less stupid.
That is also why I’m so focused on protocols, contracts, hand-offs and trust models. Without those, you just have a room full of very enthusiastic interns shouting at each other. With them, you start to get something closer to coordinated work.
One thing that does make me laugh, though, is that even in my synthetic organisation, I have somehow recreated the class politics of enterprise software.
All the senior synthetic people get proper names.
Henry is the architect. Margo is the red team lead. They have personalities. Roles. Gravitas. Opinions. Probably opinions about wine.
And then, down in the implementation mines, all the poor software engineers are just called Mini-Claude.
Which Mini-Claude?
Who cares?
They’ve got GUIDs against their names. That’s plenty. Get back to work.
So yes, Relay may be synthetic, but apparently I have still managed to reproduce the sacred corporate hierarchy: the architects get names, mythology and strategic intent. The engineers get tickets, identifiers and a vague sense that somebody upstairs has had a “vision”.
Still slightly mad, of course.
But then again, most useful software architecture is just madness with interfaces.