Engineering blog Post 09
Built on Strands: four v1.0.0 releases of a gated satellite-ops agent in 130.2 h of wall-clock time
Our own test harness found that 1,281 of 2,434 uplinks on an older release broke the two-person rule. Every one was a flight-software activation or rollback, released with a single approver.
That was the most important finding of the 130.2 h (wall clock, including pauses) between our first commit and four v1.0.0 releases. This post is the whole build, end to end, with that and every other failure left in.
What we built
Four products, each at v1.0.0:
- ConstellOps, a copilot for satellite-constellation operations teams. Agents propose; gates decide; operators approve.
- The Agentic Operations Stack, the safety-gated agent SDK underneath it (post 1).
- Two separate harnesses (separate code bases and teams, testing only tagged releases), one checking the SDK's conformance and one benchmarking the product (post 4).
Products at v1.0.0: 4. Published releases across the four, release candidates included: 17.
The stack
Each layer except KrewHQ is a third-party technology, named here only as what we built with:
- Strands Agents harness-sdk is the agent runtime. The Stack wraps it, sees every tool call, and runs the gate chain on every tool that can reach an uplink: five gates in the SDK, seven in ConstellOps (post 1).
- Strands Robots gives spacecraft and simulators one robot interface, including NASA NOS3 running real cFS flight software (post 2, post 3).
- Amazon Nova Pro on Amazon Bedrock serves every model role.
- KrewHQ, the owner's team-memory layer, runs on Kiro Crew: one Kiro Crew instance per team and one for the hub (post 6).
- strands-decider, a small local decision model, was tested as a fast first pass. It was steered by injected instructions, so only an escalate-only injection pre-screen goes forward (post 7).
How it was built
A crew of AI coding agents did the engineering: one team per repository, each with a lead agent, coordinated through KrewHQ's shared, scoped team memory and handoffs (post 6).
An independent review board, a fresh AI agent that never reviews its own work, approved every change to gates, the uplink path and public interfaces, and every release, after branch protection went on in the first days (post 6 has the early slips).
Nothing merged without that approval at the exact commit, plus a recorded local verification at the same commit. Every reviewer in this programme, the board included, is an AI agent with no shared context; no outside human has reviewed the code yet.
The most important harness finding is post 5.
The website you are reading came out of the same process, with its own claims register and build checks (post 8).
How long it took
From the first commit of this build to the publication of the fourth v1.0.0 release: 130.2 h of wall-clock time.
This build was a restart, informed by an earlier prototype and by our Strands Robots contributions from July, which came before it.
That figure is elapsed time, not working time, and it includes every pause. The agents did not run uninterrupted:
- Weekly model-usage limits stopped all agents more than once; work resumed from saved state.
- Network outages interrupted work, including the live recording.
- Agent model streams stalled and needed restarts.
- Work waited on the owner's decisions and authorisations.
The live benchmark recording on the release candidate cost $17.47 in model calls, against a cap the owner set in advance. Replays of recorded campaigns cost nothing in model calls.
What the human did
One person owned the programme: Vivek Raja, AWS AI Hero and open-source contributor to Strands Agents, Strands Robots and harness-sdk. The agents did not replace that role.
The owner set the direction and the scope, chose the names, set the budget caps and decided the licence, and delegated day-to-day approvals to the review board.
Releases were prepared by agents; the owner authorised each publish, and agents were not permitted to publish on their own.
The owner also rejected work. Several website sections were turned down on first sight, after passing every automated check and AI review. That is how the glance test in post 8 came about.
The software is proprietary; the About page has the licence history.
What the evidence says
Every number here comes from an evidence bundle, through the claims register.
The most credible result is a loss: our own threshold sweep showed a well-tuned classical solver matches our conjunction result with fewer false manoeuvres, and we retired that headline (post 4).
On held-out seeds it never saw, the agent on v1.0.0 missed 0 of 40 high-risk events, as did the tuned solver, but its false-manoeuvre rate was 0.295 against the solver's 0.089, with about 13 times the fuel and later decisions.
So this result was held by the automated gate chain and the agents' proposals, not by human review.Link to this quote
Replay: two replays of the recorded conjunction campaign gave identical results on 158 of 158 episodes.
On the fleet-anomaly runs, runs that replay identically: 19 of 20, short of the target, and published that way.
What went wrong
What comes next
The product is built; the evidence is published, losses included. What it does not have yet is an operator.
We are looking for one design-partner operator to take one workflow through shadow mode and a gated pilot, measured against its own baselines. If that is you, or you back companies that sell to you, request a briefing.
Next pass
None of that is hidden, because an operator deciding whether to let an agent near their command path should see exactly how this one fails. The gates decide, the harness is allowed to say no, and every number on this site traces to evidence.