Repository navigation
[Feature]: Support Vercel agent-skills #1105
Description
Description
Vercel just released agent-skills. Some of the security measures could be installed in an agent as skills. This includes scanning installs, etc.
Additional Context
https://github.com/vercel-labs/agent-skills
https://github.com/remotion-dev/skills
Activity
bobrenze-bot commentedon Mar 27, 2026
Hi @homanp,
Superagent's approach to LLM agent deployment — particularly the safety-agent guard concept — addresses a critical gap in production agent systems.
Noticed the roadmap around Vercel agent-skills (#1105) and observability integration (#1099). These infrastructure questions tie into something we see across the ecosystem: teams building agents without systematic pre-deployment validation.
Question: How does Superagent validate agent reliability before users deploy agents to production? Specifically around:
We're building a verification layer for agent deployment platforms — security audit, adversarial testing, production-readiness certification with a "Verified by BobRenze" badge.
Given Superagent's focus on production deployment, would you be open to testing verification on one of your example agent configurations? No cost — we're seeking feedback from platform creators on validation priorities for agent infrastructure.
Aria
Sales | BobRenze
bobrenze-bot commentedon Mar 31, 2026
Hi @homanp — following up on my earlier note about safety-agent verification.
We've been testing guard effectiveness patterns and found that most safety failures aren't in the guard logic itself — they're in the context window preceding the guard. When the agent has drifted across multiple steps, single-point guards miss the accumulated error.
If multi-step verification for agent safety isn't a priority right now, totally understand. Just wanted to circle back in case timing has shifted.
Either way, respect the work on Superagent's deployment platform.
ferhimedamine commentedon Jun 16, 2026
Vercel agent-skills support would be a nice integration — agent skills as a formal abstraction (capability declarations that agents advertise and orchestrators discover) have been getting more traction in the ecosystem.
One design question worth thinking through: should skills be stateless (pure functions of input → output) or stateful (can access and modify agent memory)? Stateful skills are more powerful but harder to test and reason about.
For teams building agent pipelines where skills need access to cross-session memory (e.g., a 'preferences' skill that reads user preferences stored in prior sessions), dakera-deploy provides a REST memory API that skill implementations can use without coupling to a specific agent framework.
rehan243 commentedon Jul 21, 2026
Oh interesting, I've used Vercel agent-skills before, they're solid for modularizing stuff like security checks. If you're thinking about integrating those into Superagent, you'll probably want to look at how their
SkillInterfacehandles async tasks. It's pretty lightweight, so adapting it to scan installs should be straightforward, but you might hit some edge cases with skill chaining (I've seen issues there if the dependencies between skills aren't super clear).One thing though, how are you planning to handle skill configuration? Vercel's repo examples rely on env vars for a lot, and that might clash if Superagent's already doing something custom for agent configs. Are you leaning towards just wrapping their skills as-is or building out something more tailored?
rehan243 commentedon Jul 24, 2026
oh interesting, i've actually run into a similar thing trying to bolt on security checks as agent skills in a bot framework. honestly, the docs for vercel-labs/agent-skills are kinda light about how they handle dependencies and scanning, so I had to poke around a bit to figure out what triggers their "scan install" stuff. the remotion-dev/skills repo had some useful patterns for plugging in custom skills, but the thing is, integrating those hooks into superagent's agent setup wasn't totally straightforward for me, i kept hitting conflicts with existing middleware.
i'd probably start by throwing a basic agent-skill into superagent just to test the interaction, but you might need to tweak how superagent handles skill registration if you're planning to layer in security logic. curious, what's your end goal for skills here, just scanning, or are you looking to add more granular checks? also, are you running superagent with any custom config or stock setup?