Love seeing this architecture laid out clearly. The Living Map layer as a shared state between agents is a smart design — avoids the coordination overhead that kills most multi-agent systems. The challenge at scale is usually keeping that map fresh without it becoming a bottleneck. Curious how you handle conflicts when multiple agents update it concurrently.
They do coordinate in the timing, they aren’t running simultaneously. And in case of issues, the QA detects it and then I can look at the reported issues/flags and fix manually the ones that QA can’t or isn’t sure.
This isn’t a critical mission system (think payments), so the potential issues aren’t a big deal. A potential discrepancy can live in the living map for a couple of days (think of a duplicated case or tool in the living map) and it’s not the end of the world
Excellent write-up. What I liked most is that the system avoids the usual “fully autonomous agent” fantasy and instead treats agents as constrained workers operating around structured workflows, verification, and human-controlled feedback loops.
A lot of people underestimate how quickly agentic systems become unreliable once taxonomy, extraction quality, or memory layers start drifting silently in production. This article actually acknowledges those operational realities instead of pretending that prompting alone solves them.
Also appreciated the emphasis on iterative refinement over one-shot orchestration. That’s much closer to how real production systems evolve.
I write a lot about distributed systems, DDD, and production AI infrastructure myself, and this was absolutely worth the subscription.
I agree, everyone thinks about fully automated AI. Then they question its viability as LLMs aren’t predictable. Then they complain it erased their prod db. And so on.
Truth is, they are amazing tools, if you can control them. There needs to be humans overseeing systems until they can be improved. And the cycle continues.
This kinds of systems, where the heavy lifting is outsourced to agents is already 90% better than what was possible ~3-4 years ago.
The "erased their prod db" arc is unfortunately a real pattern, not a joke. The teams that skip the human oversight layer usually discover why it exists the hard way. The 90% improvement is real, but it's 90% of a job that still needs the last 10% done by someone who can be held accountable.
Love seeing this architecture laid out clearly. The Living Map layer as a shared state between agents is a smart design — avoids the coordination overhead that kills most multi-agent systems. The challenge at scale is usually keeping that map fresh without it becoming a bottleneck. Curious how you handle conflicts when multiple agents update it concurrently.
That’s a good point.
They do coordinate in the timing, they aren’t running simultaneously. And in case of issues, the QA detects it and then I can look at the reported issues/flags and fix manually the ones that QA can’t or isn’t sure.
This isn’t a critical mission system (think payments), so the potential issues aren’t a big deal. A potential discrepancy can live in the living map for a couple of days (think of a duplicated case or tool in the living map) and it’s not the end of the world
Excellent write-up. What I liked most is that the system avoids the usual “fully autonomous agent” fantasy and instead treats agents as constrained workers operating around structured workflows, verification, and human-controlled feedback loops.
A lot of people underestimate how quickly agentic systems become unreliable once taxonomy, extraction quality, or memory layers start drifting silently in production. This article actually acknowledges those operational realities instead of pretending that prompting alone solves them.
Also appreciated the emphasis on iterative refinement over one-shot orchestration. That’s much closer to how real production systems evolve.
I write a lot about distributed systems, DDD, and production AI infrastructure myself, and this was absolutely worth the subscription.
Great to hear that Alireza!
I agree, everyone thinks about fully automated AI. Then they question its viability as LLMs aren’t predictable. Then they complain it erased their prod db. And so on.
Truth is, they are amazing tools, if you can control them. There needs to be humans overseeing systems until they can be improved. And the cycle continues.
This kinds of systems, where the heavy lifting is outsourced to agents is already 90% better than what was possible ~3-4 years ago.
The "erased their prod db" arc is unfortunately a real pattern, not a joke. The teams that skip the human oversight layer usually discover why it exists the hard way. The 90% improvement is real, but it's 90% of a job that still needs the last 10% done by someone who can be held accountable.
Couldn’t agree more!