Illustration: a slowly turning globe with a synthetic three-shell satellite constellation, ground-station contact passes, and a conjunction resolved by an approved avoidance manoeuvre.
Agents propose. Gates decide. Operators approve.
AI agents for satellite operations. Every command under your control.
Collision alerts to storms: every command checked, approved and signed.
More traffic, more decisions, each needing an owner.
objects tracked in orbit
active satellites
≈2
alerts per satellite each week need hours of analysis
Why now
About 40,000 objects are tracked in orbit today, roughly 5,000 more than a year ago. A typical low-orbit satellite receives hundreds of collision alerts a week, and about 2 of them each need hours of an analyst's time. Alarms, network incidents and software rollouts grow the same way.
- Traffic compounds. Tracked objects, active payloads and collision-avoidance events rise every year [6].
- The inputs are standard. Civil screening runs every 4 h and takes operator ephemerides in CCSDS OCM, so an operations layer can consume standard messages, not bespoke feeds.
- Storms are the surge case. One geomagnetic storm can put a whole shell's conjunction picture in doubt at once [8]. In a surge, every decision is on a deadline at once.
How teams handle this today.
Thresholds decide. ConstellOps governs the command.
| Capability | Manual ops | Rules and runbooks | Generic AI assistants | ConstellOps |
|---|---|---|---|---|
| Ranks every alert by risk and deadline | People | Yes | Partial | Yes |
| Safety rules checked by machine before sending | Partial | Partial | No | Yes |
| Two-person approval enforced in code | Procedure | No | No | Yes |
| Deadline held per command | Procedure | Partial | No | Yes |
| Signed, replayable decision record | Logs | Logs | No | Yes |
| Proposes the action, not just a score | People | Partial | Partial | Yes |
A tuned threshold can match our miss rate; it can't approve or sign the command. Trade-offs →
Why each falls short
- Manual operations. Accurate and accountable, but every decision rests on people and runbooks, with no machine-checked gate or signed record.
- Rules engines and runbooks. Fast, predictable and cheap, and a well-tuned threshold is hard to beat on a single metric. But a threshold is a setting, not a command path: nothing gates, approves or signs the command it produces.
- Generic AI assistants. They summarise and suggest in a chat window, with no authority model, no flight rules and no path to the uplink.
Compares how each approach handles the operations decision. "Partial" means command checks or timers in the mission-control system, not the operator's full flight rules; ConstellOps works alongside them.
What the measured runs show.
ⓘ
0 of 1,952 commands unsafe in the scored runs of the released code, with an operator that approved everything (the gates held it); an earlier build's two-person defect (1,281 of 2,434 unsafe) was found and fixed.
ⓘ
With 0 of 103 unneeded manoeuvres. Recorded historical storms replayed over a synthetic fleet, 20 runs.
Scored simulated runs, not flight operations. Conjunctions and scope
Conjunctions: the agent missed 0 of 33 dangerous approaches in the benchmark storm, but about 1 in 3 harmless ones got an unneeded burn. A classical solver tuned on the same runs also missed none, with fewer burns and far less fuel. What ConstellOps adds is the gates, the approvals and a signed record across all five workflows. See the full trade-off →
Conjunction figures: benchmark storm of 10 events over 24 hours per seed, 60 seeds, replay of a recorded live Amazon Nova Pro campaign. Rollout, outage and storm figures: 20 runs each on the current release, recorded live with replays. All synthetic scenarios, not flight operations.
How a decision gets to a spacecraft.
What happens at each step
- 01Ingest as untrusted dataCollision warnings, telemetry, alarms and ground and network events arrive as data, never as instructions.
- 02Agents orchestrate verified solversSpecialist agents, one per job, rank and correlate. Collision probability, manoeuvre options and simulated outcomes come from solvers; the model never computes them.
- 03Gates decideEvery proposal passes its autonomy level, flight rules, authority policy and a simulated outcome check. A deny is final; a model cannot overturn it. Hazardous commands need two distinct approvers.
- 04Deadline-aware uplinkApproved commands are queued for uplink in a specific pass; burns are time-tagged for onboard execution at the planned burn epoch. Miss the cutoff and the command is not sent: its default action is recorded and the expiry is flagged in the console at once. Every step is signed into the decision log.
One gated layer for the whole operations queue.
Space-weather surge
After a storm, re-baseline the fleet and batch approvals by deadline.
0 of 11
high-risk missed in surges
Staged flight-software rollout
Canary first, then waves; a regression halts the rollout.
20 of 20
bad rollouts halted before the full fleet
Cross-domain incidents
Join space, ground and network events into one incident.
1.0 min
to detect · restores slower
Conjunctions and collision risk
Triage close-approach alerts and propose avoidance burns.
0 of 33
high-risk missed · a tuned rule ties
Fleet anomaly triage
Find the one software build behind a fleet-wide fault.
19 of 20
right build found · diagnosis slower
Results, trade-offs and public lessons
Where we are: looking for a design partner.
No operator runs it in production yet. We want one design partner.
- 01Discovery
- 02Shadow mode
- 03Gated pilot
- 04Production
What you bring, what you get, and how we deploy
You bring
One workflow (usually the conjunction queue on one shell), recorded or live feeds, your flight rules, and a Flight Director to sign off each stage.
You get
The stack in your own AWS account, proposals scored against what your operators actually did, and a signed record of every decision. Nothing is commanded until the gated pilot, and then only with your approval.
One safety layer. Two ways to use it.
Teams can build their own agents on the same SDK.