AI & agentic pipelines
Agentic development moved the install decision from a human reading a package name to a tool executing one. OrbitalReg gives that new supply chain the same discipline as the old one — curated resolution, verify-on-pull, full audit — and treats the agents' own capabilities as artifacts too.
The new risk surface
Coding assistants hallucinate package names roughly one in five
recommendations — and attackers register them. Autonomous
agents run pip install and npm install
without a human reading the name at all. We documented both
attack classes and their registry-layer containment in
the slopsquatting case study
and
the dependency-confusion case study.
The containment is the same discipline that already protects your human developers: agents resolve against your OrbitalReg, not the public internet. A hallucinated name resolves to nothing and lands in the audit trail; a freshly registered malicious package meets the pull-gate before it meets a runner. In air-gap mode, an agent's package universe is exactly what you imported — nothing else exists.
New capability
Definition
An agent skills repository is a centralized, versioned store for artificial intelligence agent capabilities that functions as a strictly governed software artifact system.
A skill changes what an agent is allowed to do — which makes an unreviewed skill the highest-leverage artifact in your stack. Teams today pass skills around as files in repos and chat threads: unversioned, unsigned, unreviewed, unauditable. That is exactly where package management stood twenty years ago, and it invites the same attacks.
OrbitalReg's agent skills repository ends that: skills are published, versioned, and pulled like every other artifact in your registry — through the same gates, under the same audit trail, inside the same air gap.
Every skill is a versioned artifact with a cryptographic signature. Agents pull a pinned, verified version — never "whatever the file happens to say today." Rollback is a version pin away.
Skills move dev → staging → production through the same promotion gates as your binaries, re-verified at each step. A skill that nobody reviewed never reaches a production agent.
Skills are scanned at upload and gated at pull like any artifact. Embedded scripts and dependencies go through the same policy engine that guards your packages.
Who published which skill, who approved it, which agent pulled it, when — append-only, filterable, exportable. Skills ship into air-gapped deployments through the same signed import path as everything else.
Running agents in your pipeline?
Rico, the founder, walks through an agent-safe registry layout: curated resolution for installs, a governed skills repository for capabilities, and what both surface in the audit trail.
