Peers & simulation
Subagents delegate. Peers are individuals.
Most multi-agent systems are one brain delegating to helpers. Our peer model — built in Antlet.OS, our peer platform — is different: peers are separate instances with their own session, their own rules and their own private background graph.
The dual-graph model
Every peer holds two graphs — think of them as two planes in depth: the view in front, the state behind it.
- Front: the interaction projection — the tidy, shareable view of a session: questions, accepted answers, next steps — computed per user, viewpoint and revision.
- Behind: the process state graph — observations, evidence, rules, decisions, conflicts and alternatives. The full case state — it never leaves the peer.
- Coordinated by construction — the front is derived from the back: an answer becomes a typed event, the owning peer computes the next state and a new projection. There is no second rulebook.
fig. 01 — the dual-graph modelfront: shared · behind: private
Two working modes — and when each is right
fig. 02 — subagents vs peersdepth card: private graph · gold dashed: simulated peer
Subagents & child agents — delegation inside one orchestrator: shared context, borrowed brain, gone when the task ends. Exactly right for splitting one job into parallel work.
Peers — independent individuals: own session, own state, own private background graph. Only projections are shared — the platform never copies a peer's internals. Right for systems that must own their decisions.
A2A + Antlet semantics — discovery, stateful tasks and artifacts run over standard A2A 1.0; our lean, versioned extension adds projections, evidence and approvals on top. Efficient on the wire, and private data never leaves its peer.
Simulation & system mocking — because a peer is a real individual behind a contract, we can swap it for a simulated one. We mock an ERP counterpart, a supplier API, a partner system or an authority and test whole journeys before go-live — and a simulated success is always labeled as simulated, never sold as real.
fig. 03 — the BITS agent layerchannels · orchestration · skills · foundation
Inside every peer
Built the same way, every time
Whether a peer or a single agent: we assemble each one from the same proven layer — swappable runtimes underneath, open protocols in between, safety by design on top. If a framework changes tomorrow, your investment survives.
- MCP-first integration — every connector built once, usable from Copilot, Claude, ChatGPT and your own agents.
- Guardrails & approvals — policies, PII redaction and human sign-off on critical actions.
- Observability — OpenTelemetry traces, cost per run and regression evals on every change.
- Memory & state — checkpointed workflows that survive restarts and can run for days.
- Compliance-aware — engineered with the EU AI Act's transparency duties and GDPR-compliant data residency in mind.
Curious what a peer for your system looks like?
Bring one integration or one counterpart system — we sketch the peer, its contract and a simulation plan in the first conversation.