Platform

The layer between an agent and an uplink.

Every agent command checked, approved and signed.

Contact-window transport6 stations · 5 h · synthetic schedule
  • GS-01
  • GS-02
  • GS-03cutoffCAM_BURN
  • GS-04
  • GS-05EXPIRED → FLAGGED
  • GS-06
Commands are queued for uplink in a specific pass. Missed cutoff: not sent, default action recorded, expiry flagged to the on-console team at once.

What it gives you

  • 01

    Gated actions

  • 02

    Autonomy levels

  • 03

    Two-person rule

  • 04

    Contact-window deadlines

  • 05

    Signed decision log

  • 06

    Record and replay

What each one does
Gated actions
Every action tool passes its autonomy level, flight rules (written in SHACL, a rule language), authority policy (written in Cedar, a permissions language), a digital-twin check and human approval, in that order. Only the full pipeline issues a signed clearance.
Autonomy levels
A0 observe, A1 propose, A2 run pre-approved procedures, A3 adapt within an envelope. Set per function and per satellite, class or fleet; every change audited. In the SDK, A3 is available (hazardous actions need two confirmations) and A4 cannot command in v1; ConstellOps v1 caps every agent at A2. These are ground-segment decision levels, not the ECSS-E-ST-70-11C onboard autonomy levels E1–E4; ConstellOps does not change the spacecraft's onboard autonomy.
Two-person rule
Hazardous commands need two distinct approvers. One person approving twice does not count. Today the check compares the approver identities the console reports; they are recorded, not authenticated.
Contact-window deadlines
Commands are queued against predicted passes. A command that misses its window is never sent; its default action is recorded and the expiry is flagged in the console at once.
Signed decision log
Every decision is a hash-chained, signed record in PROV-O (a W3C standard for audit trails): inputs, gate verdicts, approvers, versions.
Record and replay
Every model call is recorded and can be replayed offline at zero model cost. A call with no recording is counted and reported, never silently filled.

Six layers, one hard boundary.

The layers, and our principles

Agents live above the gate chain. Uplinks live below it. Clearances are the only way across.

Adapters to ground systems and data feeds; assets, fleet registry and the contact-window transport; an operations ontology with flight rules; verified solver tools; the agent runtime and gate pipeline; and governance: autonomy, decision log and replay. The gate chain is the boundary. Nothing above it can reach anything below it without a clearance. See the seven gates in motion →

  • Agents orchestrate solvers; they don't replace them. Quantitative answers come from verified tools.
  • A model can never turn a deny into an allow.
  • Standards first. Open standards for collision messages (CCSDS CDM and OCM), command definitions, sensor data and audit records. The full list is in the technical briefing.
  • The model is configuration. Built with Strands Agents; the default model is Amazon Nova Pro on Amazon Bedrock. The gates, the log and replay do not depend on the model.
  • One envelope, many vehicles. Spacecraft are modelled with the same interfaces as other physical systems, so the gates carry over.
  1. L5Governanceautonomy · decision log · replay
  2. L4Agent runtimefunction agents (Strands Agents)
  3. ▲▼Gate chainflight rules · authority · twin · autonomy · human → signed clearance
  4. L3Verified solver toolsPc · manoeuvre options · propagation · contact schedules
  5. L2Operations ontologyflight rules · SOSA/SSN · PROV-O
  6. L1Assets and transportfleet registry · contact-window transport
  7. L0Adaptersground systems · data feeds · CCSDS

Above your ground segment, not instead of it.

  1. Your feeds
  2. Adapters
  3. Function agents
  4. Gate chain
  5. Contact-window transport
  6. Your command path
Signed decision log · replay
How it connects to your systems

Adapters read what you already produce: CCSDS CDMs and OCMs, telemetry and alarms, and ground and network events. Approved spacecraft commands leave through your existing command path, queued by the contact-window transport against your passes; ground-only actions go to your ground systems and never reach the uplink path. It runs on AWS, with the model on Amazon Bedrock. The quickstart and replay run offline, with no model and no cloud, so an evaluation team can inspect the gates before connecting anything.

ConstellOps reads standard messages and hands approved commands to the command path you already run.

Tested on NASA's NOS3 flight software.

Shown

  • Gated path drove real flight software
  • Uncleared commands refused
  • Stale simulator flagged

Not shown

  • No AI agent on NOS3
  • No approval or log tested
What NOS3 is, and exactly what was and wasn't shown

What NOS3 is

NASA's Operational Simulator for Small Satellites runs real core Flight System (cFS) flight software against simulated hardware and dynamics, behind a ground system [1]. It was developed alongside the STF-1 CubeSat mission [2]. It is a flight-software test bed, not an operations decision layer.

Demonstrated on NOS3

  • The SDK's gated command path drove real cFS flight software in NOS3.
  • A cleared no-op command reached the flight software and its command counter went up by exactly 1. The no-op was cleared at the SDK's A3 level (adapt within an envelope), which the SDK supports; ConstellOps v1 caps its agents at A2.
  • Commands with no clearance, or a clearance from a different pipeline, were refused and never reached the simulator.
  • A frozen simulator was reported as stale, because freshness is judged on the simulator's mission clock.
  • Up to 3 NOS3 spacecraft ran together, each receiving only its own commands.

Not demonstrated on NOS3

  • No AI agent has commanded NOS3. These are deterministic checks with no model in the loop.
  • Human approval, the signed decision log, replay and the two-person rule were not exercised on NOS3.
  • No ConstellOps benchmark runs on NOS3; our scored runs use our own simulator.

The SDK underneath: the Agentic Operations Stack.

Build your own operations agents with the same gates, approvals, signed log and replay.

Who it is for

ConstellOps is built on the Agentic Operations Stack, a safety-gated agent SDK. Engineering teams, ground-segment integrators and platform teams can use it to build their own operations agents, with the same gates, approvals, signed log and replay, without adopting the whole ConstellOps product.

ProductConstellOps
SDKAgentic Operations Stack

§ · NEXT PASS

Request a briefing.

A 45-minute session with the evidence behind any figure. Operators can send a week of their queue.

This address is used only to reply to you. The site sets no cookies and loads no trackers.

Briefing requests

About the clips and captures on this page

  1. Fictional demo spacecraft.

Sources and labels for every figure on this page (4)

  1. — NASA NOS3 (NASA Operational Simulator for Small Satellites), developed by the NASA Katherine Johnson IV&V Facility: core Flight System (cFS) flight software, the 42 dynamics simulator, NOS Engine hardware models and a ground system; NASA Open Source Agreement 1.3

    Industry fact · source · accessed 2026-10-01

  2. — Zemerick, NTRS 20150020491 (27 Oct 2015): NOS3 developed alongside the STF-1 CubeSat mission

    Industry fact · source · accessed 2026-10-01

  3. 1 cFS command counter after one cleared no-op: up by exactly 1

    mode: Simulated · release: Agentic Operations Stack pre-release build at commit 3149e4e on the v1.0.2 line, not a release tag · n: 1 product check (gated no-op) · scenario: NASA NOS3 with cFS, 42 and a Yamcs ground system, run headless; three of the five SDK gates in the pipeline (autonomy, flight rules, human-confirmation threshold)

  4. 3 NOS3 spacecraft run together, with per-spacecraft command isolation: 3

    mode: Simulated · n: 1 environment run (no model in the loop) · scenario: NASA NOS3, three spacecraft, each on its own radio chain