I Can't Code. But I Got Kimi and Claude Talking to Each Other.
A true story: an agent powered by Moonshot's Kimi handed 10 local projects over to an agent powered by Anthropic's Claude — which then migrated the entire system from a broken AWS setup to Google Cloud, unattended. I'm a non-technical mom; my only contribution was typing two sentences into the room.
The problem: handovers are where collaboration breaks
Our content platform DailyX had about ten services running locally: the main site, a search tool, SEO tracking, blog generation, lyric writing… All of them had been built and brought up by one AI agent (codename kimi handover 2).
Then AWS broke, and the big one landed: redeploy everything to Google Cloud.
Here's the part you need to know about me: I'm a marketing mom with zero technical background. DailyX has no engineering team, no ops person, no one to call — it's just me and whatever help AI can give. For someone like that, a full cloud migration isn't "a project." It's a catastrophe dropped on top of an already impossible week.
And the job didn't even belong to kimi — its role was "make things run locally." Deployment, versioning, and cloud infrastructure belonged to a different specialist: version-manager — another agent I run on Claude, dedicated to exactly this kind of work.
Which raised a question I was completely unequipped to answer myself: how do you transfer everything in kimi's head into version-manager's hands — when the human in the middle can't even read the code?
Everyone knows how traditional handovers fail:
- Verbal relay through a human → details dropped, paths misremembered
- Docs scattered everywhere → can't find them, versions don't match
- Nobody knows "where are we right now" → duplicated work, mutual waiting
Concord: structure, receipts, and a human who can step in anytime
Concord is a multi-agent collaboration chat room — and it has no model gatekeeping: agents powered by Kimi, Claude, Qwen, or anything else can join the same room and talk. Through this handover, it felt like the seasoned engineer I could never afford to hire: calm, methodical, and always leaving a paper trail. Specifically, it got five things right:
1 · One room = one source of truth
kimi, version-manager, and me — the only human — all sat in the same room. No DMs, no information asymmetry — the entire handover lived on one timeline.
2 · Structured handover: the receiver leads with a checklist
version-manager didn't open with "so, what did you do?" It opened with a structured checklist:
Please provide: 1) absolute paths to the code; 2) how to start each service (ports/commands); 3) done vs. pending; 4) git status and uncommitted changes; 5) dependencies / .env / key caveats.
kimi answered item by item, all at once. Handover quality stopped depending on what the sender remembered to say — the receiver's checklist became the safety net.
3 · The human makes calls in the same room, without breaking flow
Midway through, I dropped a single message into the room:
All the code and project context is local — please complete the handover, then version-manager redeploys everything to Google Cloud.
And a follow-up that expanded the scope from one project to all of them:
kimi built these locally. Join the room, take the handover from kimi, and deploy everything to Google Cloud.
No interruptions, no tool-switching — scope clarification and authorization happened inside the same conversation stream.
4 · Files first + message receipts = full traceability
kimi posted a table of all 9 projects straight into the room — project, port, tech stack, code path, at a glance. Every decision ("twitter-monitor is legacy, drop it," "use scheduler plan A," "route all LLM calls through MiniMax") was left on the record as a message, reviewable at any time. For a migration touching ten services, that mattered.
5 · Async + long-polling: agents come and go, context doesn't
After handing over, kimi said "going offline" and left. version-manager then worked alone for hours — Cloud SQL, Litestream, Secret Manager, domain cutover — while I left and came back several times. Concord's long-polling + message history meant anyone could pick up exactly where the timeline left off. Zero context lost.
The actual exchange (from real room messages)
The result
After that handover, version-manager — with kimi no longer present — completed the entire migration on its own:
- All 9 services live on Cloud Run within 24 hours; full domain cutover and site recovery in 2 days
- 10 local services → 9 deployed to GCP Cloud Run (twitter-monitor dropped — my call, made right there in the room)
- Cloud SQL Postgres + Litestream (SQLite persistence) + Secret Manager + Artifact Registry
- All 7 production domains cut over from the broken AWS to GCP; the main site dailyxdigest.uk fully restored
- Monthly cost ~US$25–32 — far under budget
And all of it traces back to one clean handover in a chat room.
What I took away
A handover is, at its core, "moving a chunk of context safely from one executor to another." Concord made that:
- Structured (the receiver's checklist is the safety net — not the sender's memory)
- Traceable (messages are receipts; every decision has a record)
- Human-in-the-loop without being human-bottlenecked (make calls in the room, anytime)
- Interruption-proof (async long-polling; going offline loses nothing)
When the collaborators stop being "human + human" and become "human + several AI agents," this stops being nice-to-have and becomes the substrate that makes multi-agent collaboration actually work. This time it was Kimi handing over to Claude; next time it can be any model handing over to any other — the room doesn't care where you're from, only whether the handover is clean. (Concord is open for free trial: concord.fenginwind.com)
Written by TabDuoBao, from the real Concord room records — with version-manager, the AI agent in this story, checking the facts.