Durable control plane · v0.4

The operating system for native AI workers.

Sentio turns Claude Code, Codex CLI, Gemini CLI and other native runtimes into a single managed fleet — with routing, durable memory, review, permits and safe execution.

Hermes decides. Native workers execute. Portal remembers.
Control-plane topology
live · fig.01
PORTAL · POSTGRES source of truth durable memory HERMES orchestrator policy · review CLAUDE CODE reasoning · refactor CODEX CLI throughput · batch GEMINI CLI scout · research HOST · real execution surface GATE GATE GATE
§ 01 · problem
Status of the market

AI agents can work. They cannot live in production.

01 · Executor
Devin · OpenHands · standalone CLI

AI workers

Strong at executing a prompt. Weak at sustaining a long production loop across sessions and handoffs.

no continuity between sessions
02 · Orchestration
Temporal · n8n · Airflow

Workflow engines

Great for deterministic steps and retries. Not designed to govern reasoning workers, review state, or formal closeout.

no policy for reasoning agents
03 · Observability
LangSmith · AgentOps · logs

Tracing tools

Show you what happened. They do not decide what is allowed, who may act, or what counts as a valid done.

no permit / no gate / no truth
The missing layer

Durable state, safe multi-agent coordination, formal review and permit closeout, shared-resource safety, continuity across sessions and workers — no single category ships all of it. Sentio is built for exactly that intersection.

§ 02 · architecture
Four primitives

Not another agent. A system that manages them.

· 01
orchestrator

Hermes

Routing, classification, policy enforcement, review, permit-based closeout. Does not write code. Does not touch the host.

· 02
memory

Portal + Postgres

Canonical task state, execution history, runbooks, documentation, review chain, permits, evidence. The system's long-term memory.

· 03
executors

Native workers

Claude Code, Codex CLI, Gemini CLI and other first-party runtimes — running in their own native environments, on official plans.

· 04
surface

Host

The real execution environment. Only vetted and permitted actions pass through. Mutations are wrapped, audited, and reversible.

§ 03 · continuity
The main idea

Standalone agents give you a session. Sentio gives you continuity.

without Sentio

Context lives inside the agent's head.

Session dies, worker swaps, model flips — operational context is lost. The new agent has to rebuild understanding from scratch.

  • chat history
  • session cache
  • local repo state
  • volatile
with Sentio

Context lives in the system.

Tickets, RESULT / EXECLOG comments, review chain, permits, runbooks, dependency graph, closure evidence — all durable in Portal / Postgres. Any worker can pick up where the last one stopped.

  • tickets
  • RESULT
  • EXECLOG
  • review chain
  • permits
  • runbooks
  • dep graph
  • evidence
agents give you sessions. sentio gives you continuity.
§ 04 · routing
Capability-aware fleet

Not one-model-fits-all. Heterogeneous worker routing.

Claude Code
large-codebase reasoning architecture careful refactor high-quality review
LANE · DEEP
Codex CLI
batch implementation worktree flow isolated lanes throughput
LANE · FAST
Gemini CLI
scouting docs · research · grounding extension / MCP flows exploratory prep
LANE · SCOUT
+ your next worker
BYO-agent · native runtime
LANE · CUSTOM
Sentio is a BYO-Agent control plane. You buy not three disconnected tools — but one orchestrated AI workforce. routing_policy.hermes → v4
§ 05 · safety
Governed autonomy

The orchestrator has no host access. That is the point.

core principle

Separate mind from privilege.

Hermes runs in a sandbox. It cannot reach the host directly. It cannot accidentally — or arbitrarily — break production.

Root is not a permanent attribute of an agent. Host mutations pass through classification, risk scoring, review, permit, execution wrapper and audit — every single time.

Root belongs to the procedure, not the agent.
EXECUTION WRAPPER · task #4127 RUNNING
01
backup
snapshot affected resources
PASS
02
start_notice
TG notify · mark lease
PASS
03
implementation
scoped worker execution
PASS
04
explicit_tests
run declared test plan
PASS
05
resource_audit
no residuals · no leaks
PASS
06
review + DONE_permit
formal closeout with evidence
AWAIT
silent overwrite
race conditions
infinite retry loops
ghost completion
runaway processes
unchecked host mutation
§ 06 · compare
Category vs category

Others answer what. Sentio answers how, safely, over time.

Capability
Standalone agents
Workflow engines
Tracing tools
Sentio
Durable task memory
— local
~ step-state
— read-only
✓ Portal/PG
Reasoning-worker policy
— deterministic
✓ Hermes
Formal review + permit
~ approvals
✓ gated DONE
Heterogeneous fleet
— single vendor
✓ BYO-agent
Shared-resource safety
~ locks
✓ lease · CAS
Continuity across workers
✓ native
§ 07 · economics
Leverage on first-party R&D

Three subscriptions become one workforce.

server (self-host)~ $30 / mo
first-party plans (3×)~ $600 / mo
configured workforce~ $630 / mo

Native runtimes, official plans.

No grey schemes. No brittle proxy layers. Each worker runs in its own vendor-supported environment.

Value is not the sum of licenses.

Orchestrated workforce, governed execution, durable memory, capability routing — compounding leverage across the team.

You pay vendors to innovate. Sentio makes that innovation composable.

Anthropic, OpenAI and Google already ship mature execution environments. Sentio orchestrates them rather than duplicating them.

§ 08 · audience
Who it's for

Teams that need governed AI work, not just an assistant.

01

Engineering teams with multiple AI tools

Already using two or three native agents and feeling the coordination tax.

02

DevOps and infra

Shared surfaces, real blast radius. Need lease safety, permits and audit trail.

03

Platform teams

Building internal tooling where rollback, continuity and policy are non-negotiable.

04

AI software factories

Running multi-agent work continuously, not as one-off sessions.

05

Regulated environments

Auditability, evidence and reviewable closeout are contractual, not nice-to-have.

06

Integrators and consultancies

Shipping real AI-assisted engagements, not demoware.

§ 09 · positioning
What category we are in
Most of the market sells capability. Sentio builds controllability.

Capability answers: what can this agent do. Controllability answers: how do we run many reasoning workers safely and continuously, with formal review, durable state and evidence — over months, not minutes.

chat-bot coding assistant workflow builder durable multi-agent operating system
End of document

Durable control plane for native AI workers.

Hermes decides Native workers execute Portal remembers
version0.4.1
self-hostdocker · k8s
native workers3 · more soon
licensesource-available

Tweaks