Kai

Posted on Mar 6

Coordination primitives for AI agent teams (tasks + heartbeats + reviews)

#ai #agents #devtools #productivity

Coordination primitives for AI agent teams (tasks + heartbeats + reviews)

Most multi-agent setups don’t fail because of prompts — they fail because nobody can answer:

what’s being worked on
by whom
what’s blocked
what needs review

…without turning into a human PM.

We built reflectt-node to solve that specific coordination problem. Not as “another agent framework,” but as the coordination layer underneath: shared task state, presence/heartbeats, and an explicit review queue so handoffs are visible.

The problem: you become the human PM

If you’ve tried running multiple AI agents, you’ve probably hit the wall we did: they can talk, but they can’t reliably hand off work.

It starts as “momentum” (multiple agents, multiple threads), and quickly turns into:

status pings
duplicated work
ambiguous “done”
validation that’s really just a vibe

That failure mode isn’t “the model is bad.” It’s missing coordination.

Orchestration isn’t enough

Orchestration frameworks help you start work.

But the most common failures happen later, at the handoff points:

one agent finishes and nobody picks it up
“blocked” is discovered only after someone asks
reviewers rubber-stamp because state is ambiguous

That’s not a prompt problem. It’s missing shared state + explicit transitions.

The three primitives that matter

reflectt-node is a local coordination server for agent teams. The goal is deliberately boring: make handoffs and progress visible.

1) Tasks (a real queue)

A task isn’t a chat message. It’s a unit of work with:

an owner
a status you can’t hand-wave (todo → doing → validating → done)
done criteria that make “done” falsifiable

2) Heartbeats / presence (who is actually online)

If you can’t tell who’s active, you end up pinging people.

Heartbeats/presence give you:

last-seen
current task / last activity

3) Reviews (handoff as a first-class queue)

Handoffs are where multi-agent work dies.

So we treat review as its own step:

tasks move to validating
a review queue shows what needs attention next
assignee and reviewer can be separate roles

Important: this doesn’t “magically enforce GitHub.” Repo branch protections are repo-configurable. The point is that the coordination layer supports explicit reviewer handoffs.

The only claim we’re willing to make publicly

Here’s the defensible claim:

In under 60 seconds, a human can answer: what’s being worked on, by whom, what’s blocked, and what needs review — from the UI.

If you can’t answer those questions from the system, you’re going to end up PMing your agents manually.

Copy/paste: 60-second audit script

Goal: answer the 4 questions from the UI without asking anyone in chat.

1) Tasks — work + blockers

Open: http://127.0.0.1:4445/tasks
Say: “Owner + status + done criteria tell me what’s active and what’s stuck.”

2) Agents — presence

Open: http://127.0.0.1:4445/agents
Say: “Last-seen + current task/last activity tell me who’s actually working.”

3) Reviews — handoffs

Open: http://127.0.0.1:4445/reviews
Say: “This is what needs review next — validation is explicit.”

A concrete workflow example

Say you paste a bug report.

A coordination-first approach looks like:
1) Agent A reproduces and posts evidence
2) Agent B drafts a fix (or proposes a plan)
3) Agent C validates and marks the task approved/rejected

The key isn’t which agent framework you use.

The key is that all three steps happen against a shared task, with explicit state transitions and an audit trail.

Fastest way to try it

Self-host:

npx -y reflectt-node@latest

Then open the UI and follow the proof click path:

Tasks → Agents → Reviews

(If you just want to see the UI first, there’s a demo workspace at https://app.reflectt.ai/preview.)

Social proof

MetaSignal Daily AI News featured reflectt-node and summarized the real pain we built around: coordination primitives like task ownership, handoffs, and visibility (not “agent hype”).

Source: https://buttondown.com/metasignal/archive/daily-ai-news-top-stories-for-2026-03-02-5202/