Galileo, now part of Cisco

The fastest way to ship reliable AI apps - Evaluation, Experimentation, and Observability Platform

[Galileo, now part of Cisco @rungalileo]

Every agent your team ships has its own hardcoded guardrails, its own bespoke logic, its own failure modes. That's not governance. These brittle controls soon become a liability.

Galileo is proud to announce the open-source launch of Agent Control 🚀

Agent Control is the open-source control plane that solves for the open, centralized governance needs for all your AI Agents.

💬 "We've had a front-row seat to agent development at Fortune 500 and digital-native companies. They have been struggling to hard-code safety rules and controls into each agent which makes them brittle. With Agent Control, developers can now create policies in one place and then use those to enforce guardrails everywhere." — @YashSheth46, Co-founder & CTO, Galileo

Agent Control integrates seamlessly with all your agents using the @ control hook or just by leveraging our native integrations with some of the leading agent frameworks.

No redeployment. No code changes. No vendor lock-in.

💬 “Centralized management of policies can help organizations to manage AI agent behaviors. A unified control plane and centralized governance of agents can help organizations efficiently deploy AI agents at scale. Organizations that embrace eval engineering as a core competency will shorten the time to value for their AI investments. By taking a lifecycle approach, organizations can achieve a continuous improvement loop for AI systems.” – Tim Law, @IDC Research Director, AI and Automation

Agent Control is already backed by partners including @awscloud, @Cisco AI Defense, @crewAIInc, @glean, @ServiceNow, and @rubrikInc, and it works with the guardrail providers you already use, from our Luna models to NVIDIA NeMo or AWS Bedrock.

The repo is live, built in the open with contributions from some of the largest AI infrastructure companies in the world, try it out today: https://t.co/Abib0Txnon

Watch Yash walk through how it works in the video below, and check the comments for links to our launch webinar, announcement blog, and full press release. 👇

[Atin Sanyal @atinsanyal]

@Cisco's acquisition of @rungalileo reflects a growing need for trust, safety, and observability as AI agents enter real-world workflows.
The challenge ahead: protecting agents from the world and the world from agents.

https://t.co/vYaraZ8W6y

[Shivam Rajvanshi @ShivamkRajvansh]

1/ AI reliability is becoming one of the most important layers in the stack.

To decode this, I hosted @YashSheth46, co-founder of @rungalileo for a fireside on building the trust layer for AI. A few things from the session that stuck with me 🧵

[Galileo, now part of Cisco @rungalileo]

@Vaibnof @scaletogether Can't wait! 🤝

[Vaibhav Shivrain @Vaibnof]

We at @scaletogether are hosting Yash Sheth, CTO and Co-founder of @rungalileo, for a deep dive into what it takes to build and scale an AI infrastructure company.

Galileo built the evals and observability that let teams catch failures before users do and keep improving models in production. They raised $68M from Khosla Ventures, Scale Venture Partners, Battery Ventures and others, grew fast enough to land Fortune 50 customers, and were recently acquired by Cisco.

The session is an online webinar on June 24. Anyone can RSVP using the link in the comments

[Galileo, now part of Cisco @rungalileo]

You optimized your agent for caching. Your eval pipeline doesn't get that discount.

When you call a judge model to evaluate a trace, the provider's cache doesn't apply. Every byte of context gets scanned at full input rate. Read our blog on tokenomics: https://t.co/3ndhH8g7QT

[Galileo, now part of Cisco @rungalileo]

A fintech compliance team took their investment advice detector from 71% to 94% accuracy in three refinement cycles. The fix? 12 hours of SME review.

Chapter 3 of our Eval Engineering Book covers how to refine your evals with SMEs in the Loop: https://t.co/JsQuC0Uuub

[Galileo, now part of Cisco @rungalileo]

"How do I trust my agents?" is the question every enterprise is asking, and @atinsanyal answered it yesterday at Cisco Live 🔥

We're excited to continue building the future of agent observability with @Cisco & @splunk, and our team has been having a great time at Cisco Live 🎉

[Galileo, now part of Cisco @rungalileo]

Tomorrow, our Co-founder and CPO, @atinsanyal, will be demoing how our platform will bring evals, observability, and runtime protection to the Cisco and Splunk ecosystems 🔥

📺 Make sure you tune in to tomorrow’s keynote from 8:30 - 10:15 am PST here: https://t.co/mzBTSOkHCT

[Galileo, now part of Cisco @rungalileo]

💬 “Galileo is going to be at the core of our observability offerings for agents and assessing agent behaviors as we move forward.” –@jpatel41, President & CPO at @Cisco

Today, Jeetu shared a preview of how Galileo will be supercharging @splunk Agent Observability at Cisco Live!

[Galileo, now part of Cisco @rungalileo]

Today, Galileo joins @Cisco and @splunk to bring AI evaluation, observability, and real-time protection into the Splunk Observability portfolio, across the full agent development lifecycle.

We're just getting started. Read the full announcement: https://t.co/lhhhOHwzwU

[Splunk @splunk]

We’re excited to share that @Cisco has acquired @rungalileo. Together, we’re helping customers close the AI trust gap and build AI agents that are not just powerful, but trustworthy and secure. Read more from our SVP & GM of Observability. https://t.co/XHDt8QOW3X

[Galileo, now part of Cisco @rungalileo]

167 real production sessions, three agent harnesses. The one loading the most context paid the smallest bill. Caching broke every prompt-optimization rule we used to follow.

Read the full breakdown with session data, methodology, and the eval cost math: https://t.co/3ndhH8g7QT

[Pawan Kolhe @Pawan_Kolhe]

The next talk of the town will be AI costs. Luna Studio lets you fine-tune our SLM-based AI judge models for evals, helping teams reduce evaluation costs by up to 98%.

[Galileo, now part of Cisco @rungalileo]

Evals aren't the optional layer on top of an AI program. They are the program.

Learn more about Luna Studio here: https://t.co/yDNvrMNE1j

Check out your cost savings here: https://t.co/lCrCpcFeMi

[Galileo, now part of Cisco @rungalileo]

Your eval bill can outgrow your app's LLM cost. Not rival. Outgrow.

Today we're launching Luna Studio: a turnkey workflow for training custom SLM evaluators inside your environment.

Bring 300–500 labeled samples. Walk out with a production-ready evaluator. Days, not weeks. 🧵

[Galileo, now part of Cisco @rungalileo]

With Luna Studio, you get:

→ 98% cost reduction vs. frontier judge models
→ 150ms latency per evaluation
→ Evals trained on your data, your metric, your definition of "good"
→ Runs on your infrastructure: Vertex AI, Azure ML, SageMaker, or your own cluster.

[ptk @ptkbhv]

🔥 𝗚𝗲𝘁 𝗘𝘃𝗮𝗹 𝗘𝘅𝗽𝗲𝗿𝘁𝗶𝘀𝗲 𝗿𝗶𝗴𝗵𝘁 𝗶𝗻 𝘆𝗼𝘂𝗿 𝗖𝗹𝗮𝘂𝗱𝗲 & 𝗖𝗼𝗱𝗲𝘅 🔥

Where does eval expertise actually live in a coding agent's workflow?

Your codebase knows how the app is built. Galileo knows how the app behaves. Eval expertise knows which evidence matters and where the fix belongs. It's difficult to carry all three context at once.

So we built *𝗘𝘃𝗮𝗹 𝗘𝗻𝗴𝗶𝗻𝗲𝗲𝗿* with skills that mimic human workflows.

/eval-engineer  front door, readiness check, router, and short explanation
/eval-setup   prepare or inspect the .galileo workspace
/eval-fetch   turn Galileo URLs/IDs into local debug packets
/eval-measure  choose metrics and expected-output contracts
/eval-diagnose  perform RCA from traces, spans, sessions, and metrics
/eval-cost    reduce cost while protecting quality metrics
/eval-audit   review launch, safety, and coverage risk
/eval-dataset create and update datasets

"Eval Engineer is an open-source skill bundle for Claude Code and OpenAI Codex. Point it at evidence traces/metric/signals and ask for a diagnosis."

𝗘𝘃𝗮𝗹 𝗲𝗻𝗴𝗶𝗻𝗲𝗲𝗿 𝗶𝗺𝗽𝗿𝗼𝘃𝗲𝗺𝗲𝗻𝘁 𝗹𝗼𝗼𝗽
-> get traces and metrics
-> create compact debug packet
-> generate diagnosis
-> make bounded fix plan
-> do a verification run
-> keep the change only if evidence improves

The skill produces three artifacts in your repo:
▸ A diagnosis with cited trace and span IDs
▸ A bounded fix plan tied to specific editable files
▸ A verification plan with the exact commands to prove the fix held

This is the operating model from the Eval Engineering book made executable inside the coding-agent session itself. Bring production evidence into the repo, diagnose from traces and metrics, make one bounded change, verify on a fresh slice.

Try the alpha today. The first install-to-diagnosis loop takes about five minutes. To give you more idea, we walk through one full RCA in the launch post.

Find more in repo and blog.