HandoffKit: PlatformPilot-style handoffs for AI agent systems

Built by DynG AI (https://dyng.ai), HandoffKit brings the coordination model behind PlatformPilot (https://platformpilot.ai) into Codex and developer workflows. Use it to design multi-agent systems that move work through explicit, traceable handoffs instead of shared scratchpads. The core move, borrowed from Go/CSP: agents communicate by passing messages, not by sharing memory. Each agent is an actor with private state, an addressable mailbox, and a single-owner loop. Coordination is messages, which makes multi-agent workflows easier to inspect, bound, and scale.

The pattern is language-agnostic; only the substrate changes. Pick your row:

Primitive	Go	Python (asyncio)	TypeScript	Rust (tokio)	Erlang/Elixir
Mailbox	chan	asyncio.Queue	async queue / AsyncIterable	mpsc::channel	process mailbox
Select	select {}	asyncio.wait(…, FIRST_COMPLETED)	Promise.race	tokio::select!	receive … end
Actor loop	goroutine	task / coroutine	async loop	spawned task	process (spawn)
Cancellation	ctx.Done()	CancelledError / Event	AbortSignal	CancellationToken	exit signal / monitor

The rest of this skill is substrate-independent.

Primitives

• Mailbox, a typed conduit. Unbuffered = a rendezvous (backpressure for free); buffered = a queue.
• Select, wait on several sources at once (peer message | timeout | cancellation). ALWAYS include a cancellation case so a wait can never block forever, an idle wait with no cancel path is a budget leak, not just a hang.
• Router, point-to-point delivery (address → one mailbox). Handoff and dispatch.
• Broker, point-to-many broadcast (one event → every subscriber). Pub/sub awareness. (This is the blackboard creeping back in, use only where ambient awareness is genuinely needed.)
• Join barrier, an agent that buffers inbound messages and only emits after N have arrived (wait for all dependencies), then resets.
• Quorum barrier, a fan-in agent that emits after the first N of total expected messages arrive.
• Budget, a selectable resource ceiling whose Done() channel closes on exhaustion so waits can stop before burning unbounded agent budget.
• Corpus, a shared referenced-knowledge store. The default merge is last-write-wins for single-writer keys; use a merge policy that is associative, commutative, and idempotent for true multi-writer keys.
• Compactor, turns rich working state into bounded handoff summaries plus corpus references.
• Nursery / Supervisor, structured concurrency for agent trees: depth limits, parent-child routing guards, and subtree cancellation.
• Dead letters, capture undeliverable routed messages for monitoring or tests instead of silently dropping them.
• Handoff, move a task by sending it; ownership transfers with the message and the sender stops touching it (single-writer at a time).
• Tracer, observe the coordination messages an agent saw and sent at its boundary. If all orchestration uses messages, that stream is the audit surface; hidden shared state still needs separate instrumentation.

Coordination shapes (same primitives, different wiring)

• Pipeline, dependent stages (A → B → C); each owns the task in turn. Use when step N needs step N-1's output.
• Pool, independent fan-out: one shared queue mailbox, N workers each receive from it; the queue itself gives load-balanced, exactly-once dispatch with no locks. Use for embarrassingly-parallel work.
• Broadcast, every subscriber reacts to the same event.
• Join / barrier, fan-in: wait for all dependencies, then proceed.

Be honest about the tradeoffs (do not oversell)

CSP is elegant for native concurrency because channels are cheap and lossless and processes are deterministic. LLM agents violate all three:

1. Messages are expensive and lossy, every hop compresses rich context into prose, billed in tokens. Prefer shallow topologies.
2. Agents carry huge private state (the context window) that can't be shipped over a mailbox without lossy serialization, keep shared knowledge in a corpus and pass references, not inlined memory.
3. Non-determinism, message ordering survives, behavioral guarantees don't; still verify at the receiving end.
4. Deadlock is worse, two agents waiting on each other just idle and burn budget; give every wait a cancellation path.

The synthesis: message-passing for control flow (delegation, results, cancellation, signaling); a shared corpus for knowledge, with explicit merge semantics instead of implicit shared-scratchpad races. The model that actually fits heavy, stateful agents is the actor model, not pure CSP, private durable state plus addressable mailboxes.

This is an existence proof of a known pattern (AutoGen, Akka, Ray, OpenAI handoffs all do versions of it), not a novel mechanism, recommend it for clarity and safety, not as a moat.

Reference implementation (Go)

The HandoffKit repo is one faithful instantiation, in Go, with deterministic unit tests and live LLM integration tests. The sketch/ package defines addresses, messages, mailboxes, agents, selectors, supervisors, corpora, and handoff context. The runtime/ package implements channel-backed mailboxes, selectors, routers, brokers, join/quorum agents, budgets, corpus/compaction, nursery supervision, dead letters, and tracing. For a Go project, the handoffkit-scaffold skill drops these packages and their runtime tests in. In another language, build the equivalent from the mapping table above, the design is identical, only the concurrency primitives differ.