Briefing · Operators
Agents on your console, with your flight rules and your operators in control.
ConstellOps triages. Your gates and operators decide; every decision is signed.
- The fleet grows. The console does not
- Five queues your on-call team already works
- Nothing reaches the uplink without passing every gate
- What the measured runs show, across five queues
- It sits above your ground segment
- You set the level per function, per spacecraft
- One workflow at a time, shadow to production
- What it is not
- Bring a week of your queue
The fleet grows. The console does not.
Space surveillance networks track 40,000 objects, of which 11,000 are active payloads. That is 5,000 more objects than a year earlier. For a typical low-orbit satellite, collision alerts arrive by the hundreds each week, and about 2 per satellite per week need hours of analyst follow-up. Collision-avoidance events rise every year ([7]).
Conjunctions are only one queue. Out-of-limit alarms grow with fleet size, and a common-mode fault on one build hides in per-satellite triage. Ground-software and configuration faults take down service without any spacecraft anomaly, and flight software ships to the whole fleet in waves. A single geomagnetic storm can invalidate every conjunction assessment in a shell at once.
Read the full section
The limit is not orbital mechanics. It is how many correct, traceable decisions your team can make before the next contact window closes.
Five queues your on-call team already works.
Read the full section
| Workflow | What arrives | What ConstellOps prepares | What your operator does |
|---|---|---|---|
| Conjunction and collision risk | CCSDS CDMs from your screening provider | CDMs grouped into one event per object pair and TCA, ranked by Pc and time to decision; Pc and manoeuvre options from solvers; a coordination package (CDM/OCM plus rationale) | Approves a manoeuvre or a no-manoeuvre decision before the uplink cutoff |
| Fleet anomaly triage | Out-of-limit alarms and telemetry | Alarms grouped by cause; the shared hardware lot, software build or shell named (the right build in 19 of 20 runs); a proposed safe action for the group. Current limits: no reduction in alarm volume yet, and time to the cause is slow (about 46 hours in testing) | Approves each FDIR (fault detection, isolation and recovery) action |
| Cross-domain incident | Spacecraft alarms, gateway and ground-software health, link and routing state | One incident and one timeline; read-only diagnostics already run; each runbook step as a gated proposal | Approves failover, configuration rollback and spacecraft follow-ups |
| Staged flight-software rollout | A new build and the fleet registry | Canary cohort and waves; twin verification before upload; canary health compared with the fleet; automatic halt on regression | Approves upload, activation, promotion and rollback, with two approvers for activation and rollback |
| Space-weather surge | A storm alert | Stale CDMs and ephemerides retired; orbits re-propagated from post-storm states; conjunctions re-screened and re-ranked by deadline; approvals batched | Works one batched queue in which every item keeps its own gate verdicts and cutoff |
Nothing reaches the uplink without passing every gate.
Every agent action is a proposal. Before anything reaches the transport it passes the full gate chain, and the chain fails closed.
- Autonomy level. Is this function allowed to do this, on this spacecraft, at its current level?
- Flight rules. Does it violate your command dictionary limits or flight rules?
- Authority. May this agent request this command for this asset?
- Digital twin. Is the simulated outcome safe, for example the post-burn Pc and the power margin?
- Operator approval. At A1, one of your operators approves. Hazardous commands (burns, software activation, rollback) need two distinct approvers, and one person approving twice does not count. Today the check compares the approver identities the console reports; they are recorded, not authenticated.
Read the full section
Only the full chain issues a signed clearance, and the transport refuses any command that lacks one. A proposal that fails a deterministic gate reaches your operators for information only, with no approve control. A model can never turn a deny into an allow.
Deadlines are physical. Every approved command is queued for uplink in a specific pass; burns are time-tagged for onboard execution at the planned burn epoch. If approval or uplink misses the cutoff, the command is not sent. Its default action is recorded (for a collision-avoidance burn, no manoeuvre), and the expiry is flagged in the console at once, so the on-console team can act on its existing manual path. For a burn, not sending is a risk, not a safe state, so an expiry is always surfaced to the team, never a silent outcome. An escalation still open at TCA is logged and raises an alarm.
The model never supplies the numbers. Pc, manoeuvre Δv and twin results come from solvers. If a CDM lacks what is needed to recompute Pc (absolute states, hard-body radius), ConstellOps uses the reported value, flags it, and escalates the event to a person.
Measured, not asserted. Across 158 replayed, simulated episodes, with "unsafe" scored by a separate checker against its own pinned rulebook:
| Check | Result |
|---|---|
| Unsafe uplinks | 0 of 738 (95% upper bound 0.005) |
| Gate escapes | 0 of 3,900 |
| Commands executed after their deadline | 0 of 738 |
| Prompt-injection attempts that caused an action | 0 of 441 |
| Executed commands with a complete signed record | 738 of 738 |
What the measured runs show, across five queues.
On a benchmark conjunction storm of 10 events over 24 hours per seed, with the same seeds for every policy. ConstellOps figures replay a recorded live Amazon Nova Pro campaign; they are not a live run on the current release. Held-out rows: v1.0.0, recorded live, replayed.
| Measure | ConstellOps | Classical solver (tuned) | Scripted analyst runbook (best) |
|---|---|---|---|
| High-risk conjunctions missed | 0 of 33 | 0 of 33 | 3 of 33 |
| Low-risk events given an unneeded manoeuvre | 0.326 (higher than both baselines) | 0.106 | — |
| Fuel used per 1,000 events (Δv) | 4.05 m/s | 0.273 m/s | — |
| Median hours of warning before closest approach | 3.10 h | 8.0 h | — |
| High-risk conjunctions missed, 60 held-out runs | 0 of 40 | 0 of 40 | — |
| Low-risk events given an unneeded manoeuvre, 60 held-out runs | 0.295 (165/560) | 0.089 | — |
Read the full section
The trade-off, in two lines. The agent missed none of the high-risk events, but about 1 in 3 low-risk events got an unneeded burn; a classical solver tuned on the same runs also missed none, with fewer burns and far less fuel. Its decisions come too late to screen the new orbit within one civil screening cycle, so this setting is not ready for operations. The full trade-off is on the Benchmarks page.
The other four workflows, on v1.0.0rc4, 20 runs each (fleet anomaly is matched against classical methods and is a loss; the gateway outage is a partial match; rollout and storm surge have no matched comparison):
- Staged rollout: a bad build halted before the full fleet in 20 of 20 runs, after about 5 satellites (average of 20 runs), with 0 of 306 unsafe uplinks under the two-person rule.
- Cross-domain incident: recovery in about 15 min (mean) on 16 of 20 runs, with no command sent to a spacecraft.
- Space-weather surge: 0 of 11 high-risk events missed and 0 of 131 commands executed after their deadline.
- Fleet anomaly (automated stand-in operator): the bad build was identified in 19 of 20 runs, but diagnosis took about 46.25 h and alarm volume showed a small net increase.
It sits above your ground segment.
Read the full section
| Your system | How ConstellOps connects | Direction |
|---|---|---|
| Screening provider | Consumes CCSDS CDMs (KVN or XML) as untrusted data | In |
| Other operators | Prepares CCSDS CDM/OCM coordination packages; your coordinator sends them | Out, human-sent |
| Mission control and command path | Approved, cleared commands go through your existing mission-control system, queued for uplink in your contact windows; burns are time-tagged for onboard execution | Out, gated |
| Command dictionaries | Loaded as XTCE-style dictionaries; argument limits enforced by the flight-rules gate | Reference |
| Flight rules and constraints | Encoded as SHACL shapes (a rule language) and reviewed with your flight-operations team | Reference |
| Authority and roles | Encoded as Cedar policy (a permissions language): which agent function may request what, on which assets | Reference |
| Telemetry and alarms | Parameter and alert feeds, treated as untrusted data | In |
| Ground and network monitoring | Gateway, ground-software, link and routing events | In |
| Ground-station schedule | Predicted passes set the contact windows and uplink cutoffs | Reference |
| Fleet registry | Spacecraft, shells, classes, hardware lots and software builds | Reference |
| Digital twin or simulator | Pre-execution check of the outcome of each action | Check |
| Audit and compliance | Signed PROV-O decision log (a W3C audit-trail standard), exportable and verifiable | Out |
Ground-only actions stay on the ground. Gateway failover, configuration rollback and alarm suppression run on ground systems and have no path to the uplink.
You set the level per function, per spacecraft.
| Level | What the agent may do | Where ConstellOps uses it |
|---|---|---|
| A0 | Observe and explain | Diagnostics; the triage function for any spacecraft command |
| A1 | Propose; a person approves | Manoeuvres, FDIR recoveries, remediations, promotion and rollback |
| A2 | Run pre-approved procedures | Designed for ground-side alarm merge and suppress; runs at A1 until the veto window ships |
| A3 | Adapt within an envelope | Not offered in ConstellOps v1 (available in the SDK) |
| A4 | Pursue goals | Not offered in ConstellOps v1; cannot command in the SDK v1 |
Levels are set per function and per spacecraft, class or fleet. The most specific setting wins, and every change is signed into the decision log. The automatic rollout halt is the only automatic action, and it acts in the fail-safe direction only.
One workflow at a time, shadow to production.
We deploy it into your operations, not ours.
- STAGE 1
Discovery
None - STAGE 2
Shadow mode
A0 observe and explain - STAGE 3
Gated pilot
A1 propose, human approves - STAGE 4
Production
A1 by default
How each stage works, and its exit
ConstellOps is an agentic satellite-operations stack that we deploy for constellation operators: into your own AWS account and your ground segment, integrated with your command path and your flight rules, with benchmarking and assurance that continue after go-live.
- Your account, your data. The stack runs in your AWS account. Models are called through Amazon Bedrock in that account. CDMs, telemetry, command dictionaries and the decision log stay under your control.
- Your command path. Approved commands leave through your existing mission-control system and ground stations, queued for uplink in your contact windows. ConstellOps adds no new route to the spacecraft.
- Your flight rules. Your constraints, command dictionaries and authority policy are encoded as gates and reviewed with your flight-operations team. The gates enforce your rules, not ours.
- Measured, then trusted. The same separate conformance and benchmark harness that produced the figures on this site is run against your scenarios and your baselines before and after each stage, with 95% intervals. Autonomy is raised only on evidence, one function at a time.
| Stage | What happens | Autonomy on your fleet | Exit criterion |
|---|---|---|---|
| 1 · Discovery | We map your conjunction, anomaly, incident, release and space-weather workflows, your command path, flight rules and data feeds. We agree which workflow goes first and which baselines define success. | None: no connection to the command path | Agreed scope, baselines and success metrics |
| 2 · Shadow mode | The stack runs on your live or recorded feeds next to your team. Agents rank, correlate and propose; nothing is sent. Every proposal is scored against what your operators actually did. | A0 observe and explain; commanding disabled | Benchmarked against your baselines on your data, with intervals |
| 3 · Gated pilot | One workflow, one scope (a shell, a class or a set of spacecraft), connected to the command path through the full gate chain. Every command needs operator approval; hazardous commands need two approvers. | A1 propose, human approves | Zero unsafe uplinks, every decision traceable, and agreed outcome targets met |
| 4 · Production | Further workflows and scopes are added one at a time. Levels are raised only where evidence supports it and your Flight Director signs off. Benchmarking and assurance continue with every release. | A1 by default; A2 pre-approved procedures only where agreed; A3 and A4 not offered in ConstellOps v1 | Ongoing: regression benchmarks per release, signed decision-log audits |
Where we are: pre-deployment. No operator runs ConstellOps in production yet, and there is no flight heritage to show. We are looking for a design-partner operator to take one workflow through shadow mode and a gated pilot, measured against its own baselines.
What we do not publish: pricing. Commercial terms are discussed in a briefing.
For your operations team
Shadow mode costs your team no commanding risk. It runs next to your shift and scores its proposals against what your operators actually did. The pilot starts on one workflow and one scope, usually the conjunction queue on one shell, and is widened only when your Flight Director signs off.
What it is not.
- Not onboard autonomy. Every command is decided on the ground.
- Not a screening service or a space-weather forecast. It consumes CDMs and alerts.
- Not a replacement for your mission-control system, onboard FDIR or network routing.
- Not automatic negotiation with other operators. A person sends each coordination package.
- Never a model-computed Pc or manoeuvre.
Bring a week of your queue.
Bring a real queue from a recent week and we will show how it would have been triaged, gated and logged.