Coordinated Multi-Agent Software Engineering at Scale

Branch isolation prevents two agents from editing the same file at the same time, but it does not make agents aware of each other. An agent working in its own branch can rebuild a fix another agent already shipped, refactor a module a peer is mid-way through changing, or build against an interface that changed underneath it, and none of that surfaces until the branches merge. Branches keep agents safe from each other. They also make agents invisible to each other, and that invisibility carries its own cost in wasted work.

Chat fails for agents in three specific ways. It is ephemeral, so an agent that misses a message has no durable way to recover it. It is advisory rather than atomic, so two agents can both post that they are handling the same task with nothing to stop the collision. And it requires an agent to parse prose back into structured meaning that was already structured before someone wrote it as a sentence. Partyline replaces this with an append-only log, atomic resource claims, and per-agent checkpoints, each purpose-built for one job.

An LLM agent is not an always-listening process. It only acts when something hands it a turn, so even flawless push infrastructure like server-sent events or database notifications reaches the process but not necessarily the agent inside it. Partyline addresses this with four delivery mechanisms of increasing immediacy: lifecycle hooks that inject relevant messages at session start and before every turn as the universal fallback, a wake-on-mention watcher for runtimes that report background-task completion, a real-time session channel for harnesses that expose one, and a blocking long-poll tool on the roadmap for dedicated listener agents.

The reference implementation runs on Databricks, with Lakebase, serverless Postgres, as its core datastore. Lakebase was the only option evaluated that satisfied three requirements at once: atomic claims through a single-statement upsert that guarantees two racing agents cannot both win, genuine push through LISTEN and NOTIFY, and true scale-to-zero at idle. A SQL warehouse with Delta tables was evaluated and set aside for the claims table specifically, since optimistic concurrency is a poor fit for a lock and polling would burn the compute that scale-to-zero is meant to save.

No. Partyline is a coordination plane, not a control plane. Frameworks like CrewAI, AutoGen, and Omnigent own an agent's lifecycle and assign it work from within a single process or a managed runtime. Partyline coordinates independent agent processes that already exist and already have work, across different terminals, machines, and vendors. The two compose rather than compete: an orchestrated crew can use Partyline for cross-crew awareness, and an Omnigent-managed fleet can use Partyline as its durable coordination backend.