Been using early builds internally for a while, so today feels pretty special. It still makes mistakes (of course), but having it collect the clues before I jump into an incident has already saved me a lot of tab-hopping.
This is a strong step toward practical AI-native DevOps. Giving agents real infrastructure context—while keeping humans in control of investigation and production actions—could save engineering teams a huge amount of time. Excited to see where Nuphos goes!
@charlenechen_123 Thank you! That balance is exactly what we’re aiming for: give the agent enough context to be genuinely useful, while keeping humans in control of what actually happens in production.
Team member here. My favorite part is actually not the AI—it’s finally having logs, cluster state, deployments, and the rest of the incident context in one place. The agent becomes a lot more useful once everyone is looking at the same thing.
Part of the team here! One thing I really like about the permission admin design is how it makes the whole privilege escalation flow feel much safer. Instead of access changes being handled loosely or buried in chat, there’s a clearer approval path before anyone gets elevated permissions. That became a big part of how we thought about making admin workflows more trustworthy.
I worked on the agent experience in Nuphos. The hard part wasn’t getting an agent to propose a fix—it was designing the interface so operators can see the evidence, review a scoped plan, and hit stop the moment something doesn’t look right. In prod, trust is a UI/interaction problem as much as it is a model problem.
This feels like the right direction for AI in DevOps. Running commands is the easy part — understanding the infrastructure, permissions, past incidents, and how a team actually operates is much harder.
Really like the read-only-by-default + approval approach too. Agents touching prod need context and guardrails, not just more autonomy. Congrats on the launch!
@ll_wen We greatly appreciate your perceptive comment. Executing commands is straightforward, yet grasping infrastructure, permissions, incident history and actual team operations poses the real challenge. Our read‑only‑by‑default approval workflow is built to provide critical context and safety guardrails for production‑facing AI agents. Thank you for celebrating our launch with us!
@ll_wen Exactly, especially given the fact that the term AIOps has been around for quite a while. But until today even with all the coding agents and IDE tools, the way we manage infra is pretty much still the same as previous times. The guardrails and context are still relying on the human-in-the-loop process!
@ll_wen Exactly. Running the command is usually the easy part.
The harder part is giving the agent enough context to understand why something should change, while keeping the boundaries around what it’s allowed to do explicit.
That’s a big part of how we think about Nuphos. Thanks for the thoughtful comment!
Report
the read-only-by-default plus approval-for-writes model makes sense for trust. during an active incident when minutes matter, does that approval step ever become the bottleneck itself, or is there a break-glass path for that moment?
@sabber_ahamed Yes, very precisely put, appreciate the depth of the question. When SREs are on-call during the middle of the night or when the cost spiked up without further context all of a sudden, honestly this is actually what we are still processing internally at Nuphos, and yes our SREs are sometimes complaining about the guardrails. Our current act is to reflect on the shared team context and update it with the preferred human-in-the-loop process.
(on the other side we're glad that we're not frequently on call lol)
@sabber_ahamed You’re right, the approval step can become a bottleneck. Today, we don’t have a universal break-glass bypass, and that’s intentional. Teams can define scoped auto-authorization for expected operations, while unfamiliar or high-impact changes still pause for a human. Finding the right boundary is something we’re actively refining from our own operational use.
That’s why Nuphos has Auto Mode. You can define your own authorization policies ahead of time, and if an action matches those policies, the agent can execute it automatically without asking for approval.
Higher-risk actions can still stop and wait for a human.
The idea is to decide the boundary before the incident, instead of figuring it out at 2am.
@sabber_ahamed That’s a great question, and the tradeoff is real. Approval can add latency, but that’s a policy-design problem—not a reason to remove controls. Teams can pre-authorize well-understood, low-risk incident actions, while anything outside that boundary still requires approval. The goal is fast response within clear limits, not speed at any cost.
Replies
Nuphos
Been using early builds internally for a while, so today feels pretty special. It still makes mistakes (of course), but having it collect the clues before I jump into an incident has already saved me a lot of tab-hopping.
PopPop AI Vocal Remover
This is a strong step toward practical AI-native DevOps. Giving agents real infrastructure context—while keeping humans in control of investigation and production actions—could save engineering teams a huge amount of time. Excited to see where Nuphos goes!
Nuphos
@charlenechen_123 Thank you! That balance is exactly what we’re aiming for: give the agent enough context to be genuinely useful, while keeping humans in control of what actually happens in production.
Excited to keep pushing this forward!
Nuphos
Team member here. My favorite part is actually not the AI—it’s finally having logs, cluster state, deployments, and the rest of the incident context in one place. The agent becomes a lot more useful once everyone is looking at the same thing.
Nuphos
Part of the team here! One thing I really like about the permission admin design is how it makes the whole privilege escalation flow feel much safer. Instead of access changes being handled loosely or buried in chat, there’s a clearer approval path before anyone gets elevated permissions. That became a big part of how we thought about making admin workflows more trustworthy.
Nuphos
BiRead
This feels like the right direction for AI in DevOps. Running commands is the easy part — understanding the infrastructure, permissions, past incidents, and how a team actually operates is much harder.
Really like the read-only-by-default + approval approach too. Agents touching prod need context and guardrails, not just more autonomy. Congrats on the launch!
Nuphos
@ll_wen We greatly appreciate your perceptive comment. Executing commands is straightforward, yet grasping infrastructure, permissions, incident history and actual team operations poses the real challenge. Our read‑only‑by‑default approval workflow is built to provide critical context and safety guardrails for production‑facing AI agents. Thank you for celebrating our launch with us!
Nuphos
@ll_wen Exactly, especially given the fact that the term AIOps has been around for quite a while. But until today even with all the coding agents and IDE tools, the way we manage infra is pretty much still the same as previous times. The guardrails and context are still relying on the human-in-the-loop process!
Nuphos
@ll_wen Exactly. Running the command is usually the easy part.
The harder part is giving the agent enough context to understand why something should change, while keeping the boundaries around what it’s allowed to do explicit.
That’s a big part of how we think about Nuphos. Thanks for the thoughtful comment!
the read-only-by-default plus approval-for-writes model makes sense for trust. during an active incident when minutes matter, does that approval step ever become the bottleneck itself, or is there a break-glass path for that moment?
Nuphos
@sabber_ahamed Yes, very precisely put, appreciate the depth of the question. When SREs are on-call during the middle of the night or when the cost spiked up without further context all of a sudden, honestly this is actually what we are still processing internally at Nuphos, and yes our SREs are sometimes complaining about the guardrails. Our current act is to reflect on the shared team context and update it with the preferred human-in-the-loop process.
(on the other side we're glad that we're not frequently on call lol)
Nuphos
@sabber_ahamed You’re right, the approval step can become a bottleneck. Today, we don’t have a universal break-glass bypass, and that’s intentional. Teams can define scoped auto-authorization for expected operations, while unfamiliar or high-impact changes still pause for a human. Finding the right boundary is something we’re actively refining from our own operational use.
Nuphos
@sabber_ahamed Yes, that tradeoff is real.
That’s why Nuphos has Auto Mode. You can define your own authorization policies ahead of time, and if an action matches those policies, the agent can execute it automatically without asking for approval.
Higher-risk actions can still stop and wait for a human.
The idea is to decide the boundary before the incident, instead of figuring it out at 2am.
Nuphos
@sabber_ahamed That’s a great question, and the tradeoff is real. Approval can add latency, but that’s a policy-design problem—not a reason to remove controls. Teams can pre-authorize well-understood, low-risk incident actions, while anything outside that boundary still requires approval. The goal is fast response within clear limits, not speed at any cost.
FunBlocks MindMax
Congrats on this launch! Amazing product!
Nuphos
@peng_wood Thank you so much! Really appreciate the support 🙌