Why Multi-Step Agent Teams Need a Transactional Control Plane
Single-turn coding assistants hold context in a temporary chat buffer. Once tasks span codebases, Linux build hosts, scheduled verifications, and multiple specialized workers, stateless chat loops run into three recurring operational failures: two workers editing the same file simultaneously, lost progress after a process restart, and silent partial side effects on retry.
ZeroOS addresses these issues by running a persistent control plane on operator-owned Linux infrastructure, storing task trees, execution leases, and artifact checksums inside PostgreSQL rather than volatile session memory.
Architectural Division Between Planning and Execution
1. Built-In CEO Calls Configured Model APIs Directly
Inside ZeroOS, the built-in CEO planner communicates directly with the LLM endpoint configured by the engineering team (including Gemini, OpenAI-compatible gateways, Anthropic Claude, or private self-hosted models). Hermes is not required for CEO planning. Task decomposition, risk gating, and review decisions run directly against operator-chosen model APIs.
2. Modular Open-Source Executors (Including Hermes)
Once the control plane registers discrete sub-tasks, modular workers—such as Hermes or containerized shell/build runners—execute scoped assignments inside isolated workspaces. Each step claims an explicit time-bounded lease in PostgreSQL alongside an idempotency key so concurrent workers cannot collide on the same repository slice.
3. Crash Recovery and Internal Staging Failover Observations
In Fidwell isolated internal staging cluster testing (November 2026), simulated worker disconnect recovery was observed at roughly 6.99 seconds failover. Actual recovery timing depends on hardware specs, network latency, and step verification bounds; this measurement reflects an internal staging run rather than an external hardware SLA.
Connecting LongMemOS over MCP
ZeroOS workers can mount LongMemOS (MemoryBox 2.0 v2.0.4) locally via Model Context Protocol (mcp.call("longmemos.search", ...)) or local HTTP APIs. Because LongMemOS stores timestamped source passages instead of lossy chat summaries—without bundling LLM weights (embedded_models=false, ~3.1 MB compressed)—teams can switch upstream inference models without re-indexing historical project decisions.
Deployment Summary for Sovereign Linux Environments
- Control state: PostgreSQL tables for jobs, leases, step state transitions, and artifact hashes
- Execution boundary: Scoped read/write policies, controlled secret paths, and human handoff pauses when verification fails twice
- Verified hardware targets for paired memory node: Windows 11
x86_64, Linuxx86_64, and ARM64RK3588(macOS validation in progress)