Civic technologist building AI tools that help systems explain themselves. Clew Suite builder, NACM 2026 speaker. I write about hackathons, product judgment, agent readiness, and learning in public.
Agreed. I didn't realize it at the time and it makes sense now. How could the project pass without the screenshots of the AWS console? I'm thankful it was such a quick fix.
Civic technologist building AI tools that help systems explain themselves. Clew Suite builder, NACM 2026 speaker. I write about hackathons, product judgment, agent readiness, and learning in public.
Civic technologist building AI tools that help systems explain themselves. Clew Suite builder, NACM 2026 speaker. I write about hackathons, product judgment, agent readiness, and learning in public.
The one-penny example makes me want one authoritative amount in the shared record, including currency and the rounding rule. I’d make the Communication agent render that value rather than calculate it again. In a payments-ledger exercise, I’d check that the quoted amount, recorded amount and refund all agree. Accepting either cent can pass a rubric while leaving the customer facing contract undecided.
Civic technologist building AI tools that help systems explain themselves. Clew Suite builder, NACM 2026 speaker. I write about hackathons, product judgment, agent readiness, and learning in public.
Thank you. One authoritative amount in the shared record, with the currency and rounding rule stored alongside it, is a strong improvement. Having the Communication agent display that number instead of recalculating it fits its job too.
Your last line adds a layer I'd build on. Accepting either cent was the right call for the rubric, and it also left the customer-facing answer undecided. Both are true, and your approach settles the second one. I'm adding "one owner for every amount, and a check that the numbers agree from quote to refund" to my PRD template for the next build that touches money.
The shift from one chatbot to several agents only pays off when handoffs are explicit. Each agent should receive a bounded artifact—facts, a plan, a patch, or a verification result—not a giant shared transcript. Otherwise the architecture adds coordination cost without adding accountability.
Civic technologist building AI tools that help systems explain themselves. Clew Suite builder, NACM 2026 speaker. I write about hackathons, product judgment, agent readiness, and learning in public.
Thank you, this matches what I saw. In Nova Concilium, each agent wrote to its own slot in one shared record instead of a shared transcript. The Refund agent read only the Inventory agent's slot, and the Communication agent read the full record to write the reply. Every write was version-checked, so agents couldn't overwrite each other.
Where your point goes further is the artifact itself. My slots held each agent's answer as prose. A bounded artifact, like a calculation with its amount, currency and rounding rule, is what would have caught the penny, because the orchestrator's math never landed in any slot. Typed handoffs instead of prose is the change I'd make on the next multi-agent build.
Nice progression — the jump from one agent to a specialist team is where the architecture starts to matter: an orchestrator keeps each specialist's tool list and context small, so prompts compete less for attention, and a safety filter on every model call (not just at the boundary) closes the loophole of one agent inheriting another's unchecked output. The shared DynamoDB state is the part most people underestimate — five agents need one source of truth or they start contradicting each other.
Civic technologist building AI tools that help systems explain themselves. Clew Suite builder, NACM 2026 speaker. I write about hackathons, product judgment, agent readiness, and learning in public.
Yes. Going through the course I was only able to see the next project in front of me. But, looking back the progression puts everything together. Agreed re DynamoDB! The agents would go off on their own and disagree without it!
Hey, I'm Hema 👋
Developer, writer, and creator of Dev Opportunity Radar, a weekly series published every Friday on DEV, helping people discover opportunities they might otherwise miss.
Civic technologist building AI tools that help systems explain themselves. Clew Suite builder, NACM 2026 speaker. I write about hackathons, product judgment, agent readiness, and learning in public.
Hey, I'm Hema 👋
Developer, writer, and creator of Dev Opportunity Radar, a weekly series published every Friday on DEV, helping people discover opportunities they might otherwise miss.
Civic technologist building AI tools that help systems explain themselves. Clew Suite builder, NACM 2026 speaker. I write about hackathons, product judgment, agent readiness, and learning in public.
A commendable journey! What stands out is not merely the leap from one chatbot to five collaborating agents, but the willingness to learn from each setback and refine the craft. That is the true spirit of engineering—progress through curiosity, persistence, and experience.
We're a place where coders share, stay up-to-date and grow their careers.
The 120 out of 120 score getting sent back is the telling part. A perfect test run still needed Console screenshots as proof.
Agreed. I didn't realize it at the time and it makes sense now. How could the project pass without the screenshots of the AWS console? I'm thankful it was such a quick fix.
Glad that one landed quickly for you. A perfect score still asking for screenshots is a strange kind of pass.
Ha. Yes, they wanted proof that the console and knowledge bases were set up in AWS.
Screenshots as receipts, then.
Exactly 💯
The one-penny example makes me want one authoritative amount in the shared record, including currency and the rounding rule. I’d make the Communication agent render that value rather than calculate it again. In a payments-ledger exercise, I’d check that the quoted amount, recorded amount and refund all agree. Accepting either cent can pass a rubric while leaving the customer facing contract undecided.
Thank you. One authoritative amount in the shared record, with the currency and rounding rule stored alongside it, is a strong improvement. Having the Communication agent display that number instead of recalculating it fits its job too.
Your last line adds a layer I'd build on. Accepting either cent was the right call for the rubric, and it also left the customer-facing answer undecided. Both are true, and your approach settles the second one. I'm adding "one owner for every amount, and a check that the numbers agree from quote to refund" to my PRD template for the next build that touches money.
The shift from one chatbot to several agents only pays off when handoffs are explicit. Each agent should receive a bounded artifact—facts, a plan, a patch, or a verification result—not a giant shared transcript. Otherwise the architecture adds coordination cost without adding accountability.
Thank you, this matches what I saw. In Nova Concilium, each agent wrote to its own slot in one shared record instead of a shared transcript. The Refund agent read only the Inventory agent's slot, and the Communication agent read the full record to write the reply. Every write was version-checked, so agents couldn't overwrite each other.
Where your point goes further is the artifact itself. My slots held each agent's answer as prose. A bounded artifact, like a calculation with its amount, currency and rounding rule, is what would have caught the penny, because the orchestrator's math never landed in any slot. Typed handoffs instead of prose is the change I'd make on the next multi-agent build.
Nice progression — the jump from one agent to a specialist team is where the architecture starts to matter: an orchestrator keeps each specialist's tool list and context small, so prompts compete less for attention, and a safety filter on every model call (not just at the boundary) closes the loophole of one agent inheriting another's unchecked output. The shared DynamoDB state is the part most people underestimate — five agents need one source of truth or they start contradicting each other.
Yes. Going through the course I was only able to see the next project in front of me. But, looking back the progression puts everything together. Agreed re DynamoDB! The agents would go off on their own and disagree without it!
Congratulations on completing the nanodegree! So happy for you 😃💙
Thank you so much! Did you finish, too? How did you like the program?
Still working on it, so hoping to finish soon! But yeah, the program is great so far.
Excited to see your take on it once you ace it!
A commendable journey! What stands out is not merely the leap from one chatbot to five collaborating agents, but the willingness to learn from each setback and refine the craft. That is the true spirit of engineering—progress through curiosity, persistence, and experience.