tmuhlestein

A lot of the conversation around AI and coding focuses on whether models will replace software engineers.

What seems to be happening in practice is slightly different.

AI is getting very good at generating code, but the hardest engineering problems were never about typing code in the first place. They’re about:

-system architecture and abstractions -understanding complex systems -reasoning about failure modes -operating systems at scale -deciding what should actually be automated

If anything, AI tools push engineers to start automating more of their own routine work, which frees time to focus on the harder system-level problems.

In that sense, the role of engineers shifts from “writing code” to designing and governing systems that increasingly include AI components.

Curious how others here see the role evolving as AI coding tools improve.

One challenge many teams are hitting right now is the gap between AI demos and production AI systems.

This AWS case study covers some of the work GoDaddy has been doing to operationalize AI workflows using AWS infrastructure.

One example highlighted is Lighthouse, a system built using Amazon Bedrock to analyze large volumes of customer support interactions and extract patterns that can improve customer experience and operational efficiency.

What’s interesting from a systems perspective isn’t just the model usage, but the broader shift toward treating AI as part of production infrastructure:

data pipelines that continuously feed models

orchestration layers around LLM workflows

recursive learning platforms that allow real-world signals to continuously improve systems

integration into existing operational systems

The biggest challenge for many organizations right now isn’t building models — it’s building reliable systems around them.

Curious how others here are approaching AI infrastructure vs experimentation.

The interesting part here is publishing agent identity to public DNS rather than keeping it in a closed registry.

This separates identity (globally resolvable) from enterprise governance and orchestration.

One consistent pattern in production AI work is that model quality is rarely the main bottleneck.

Trust, identity, security, and operational discipline tend to determine whether systems can actually scale beyond pilots—especially as agentic systems become more autonomous.

Curious how others here are thinking about governance and reliability for AI systems operating in real-world environments.

LLMs are increasingly being used as evaluators — grading outputs, ranking options, and even guiding autonomous agents. But raw model scores are unstable: they drift over time, vary across models, and don’t map linearly to real quality.

This write-up explores why LLM-based judging is unreliable out of the box, how calibration curves can improve stability, and why multi-model consensus may be necessary for trustworthy evaluation systems.

Curious how others in the industry are handling calibration, drift mitigation, or cross-model agreement for “LLM-as-a-judge” setups.

GoDaddy just published the Agent Name Service (ANS) API) and an accompanying standards site for defining identity and lifecycle management for autonomous agents.

ANS pairs a human-readable name with cryptographically verifiable identity, using PKI/X.509 and DNS-style discovery. It’s designed to work alongside frameworks like A2A, MCP, and other emerging agent protocols.

What’s new:

Public ANS API is now live (registration, discovery, certificates, revocation, etc.) ANS Standards site with the current spec + architecture Open adapter layer for interoperability without framework lock-in

GitHub repo available for builders

Links:

Standards → hhttps://www.agentnameregistry.org/ Repo → https://github.com/godaddy/ans-registry API keys → https://developer.godaddy.com/keys

Given the rise of autonomous agents, this aims to solve identity, trust, and lifecycle governance at internet scale.

Curious what HN thinks: Do we need shared, neutral identity infrastructure for agents — or should each framework solve this independently?

At GoDaddy, our engineering team has been experimenting with how to make AI agents more modular, auditable, and production-ready. Instead of treating agents as monoliths, we designed them around four core tools that separate decision from execution:

* MemoryTool – maintains persistent context and user continuity

* CompletionTool – defines task completion and termination criteria

* UserInteractionTool – handles clarifications, approvals, and feedback loops

* DelegationTool – enables handoffs between agents or humans

This approach makes agent behavior transparent, versionable, and safer to scale. We’ve open-sourced our thinking and early framework notes in a technical write-up here:

Building AI Agents at GoDaddy – An Agent’s Toolkit https://www.godaddy.com/resources/news/building-ai-agents-at...

Would love to hear feedback from others working on modular or composable agent architectures — especially how you handle agent memory persistence and versioning.

The last few months have proven that autonomous agents are testing the web’s guardrails — from robots.txt to checkout flows — and trust is the casualty.

At GoDaddy, we’re addressing this with the Agent Name Service (ANS) open standard — a framework introducing a digital passport for autonomous agents. By binding each agent’s identity to its exact version of code, ANS creates verifiable trust even in multi-agent chains, reducing the risk of compromised code or “trust rot.”

This is more than a technical milestone — it’s the foundation for secure, scalable agent-to-agent commerce. This post follows our announcement from last month: ANS is now live, and we’re partnering with select companies to roll out implementations in the near future.

ANS provides the PKI-based foundation and governance model enterprises need to trust the shift toward a decentralized, agent-driven internet — acting as the bridge between today’s Web2 enterprise reality and the decentralized agentic future.

Exciting times ahead as we continue shaping the agent economy.

Read more in Scott Courtney’s latest post: Why the Agent Economy Needs a Digital Passport https://www.godaddy.com/resources/news/why-the-agent-economy...

What do you think — could verifiable, version-bound identity be the missing piece for safe autonomous collaboration?

In most ad systems, the logic is simple: highest bid wins. But simple doesn’t mean smart.

At GoDaddy, we asked a harder question: Can we optimize both revenue and user satisfaction—without sacrificing either?

So, we built DeepAd: a deep learning ranking model that blends monetization, behavior, and intent to deliver ads that are both profitable and meaningful.

Core Concepts

• Context-aware ranking: Goes beyond “highest bid wins,” factoring in intent, engagement, and diversity. • Transfer learning from DeepRank: Leveraging billions of domain-search interactions to give our ad model a head start. • ReTiRe framework: Balancing Relevance, Time, and Revenue for a truly adaptive signal mix. • Real behavioral supervision: Learning directly from shopper actions — what they clicked, bought, or skipped.

Why this matters:

• Shoppers see domains that actually fit their intent. • Registries get fair exposure — niche top-level domain (TLD) can compete on merit, not budget. • GoDaddy achieves sustainable revenue without degrading user trust.

Key takeaway: Building intelligent ad systems isn’t just about monetization — it’s about aligning incentives between users, advertisers, and platforms through smarter modeling.

Dive into the technical deep dive: Balancing Revenue and Relevance: Optimizing TLD Ad Rankings https://www.godaddy.com/resources/news/balancing-revenue-and...

Curious to hear your perspective — share your thoughts below!

Smarter AI Starts with Consensus

The next frontier in AI isn’t about bigger models — it’s about smarter collaboration.

At GoDaddy, we’re exploring Consensus of LLMs (CoL) — an orchestration approach that brings multiple Large Language Models together to reach agreement before producing an answer.

By combining diverse perspectives from different models — GPT, Claude, Gemini, Llama, and others — CoL helps us:

• Reduce hallucinations in AI-generated outputs • Increase factual accuracy and transparency • Make AI safer for real-world, high-stakes business decisions

From customer support bots to security guidance and website content creation, this approach is making AI more reliable, explainable, and aligned with customer trust.

Read more from Jay Gowdy on how GoDaddy is pioneering trustworthy AI at scale: Consensus of LLMs: Scaling Accuracy Beyond a Single Model https://www.godaddy.com/resources/news/consensus-of-llms-sca...

Curious to hear from the AI community: could consensus between models be the key to more trustworthy systems? Share your thoughts — we’d love to continue the conversation.

Innovation doesn’t happen by accident — it happens through experimentation

Together with @Marshall Cahill, we shared how A/B testing helps us innovate at GoDaddy and turn bold ideas into a scalable impact.

In this piece, we cover how our teams:

• Turn bold ideas into safe, learnable experiments • Use Zone to Win to connect tests directly to business goals • Leverage culture and tools like Hivemind to make experimentation scalable

This mindset is shaping how we build, learn, and adapt in the era of AI — combining creativity with measurable progress.

Read more: How A/B Testing Fuels Innovation at GoDaddy https://www.godaddy.com/resources/news/how-a-b-testing-fuels...

As I mention in the blog: We don’t debate the future — we experiment our way into it. Evidence beats opinion, every time.

Curious to hear your thoughts — how is your team using experimentation to turn ideas into impact?

Travis Muhlestein

At GoDaddy, we’ve been exploring ways to make large-scale LLM integrations more efficient — not just for us, but for the entire ecosystem.

Today, most Model Context Protocol (MCP) clients inject every tool definition into every prompt. That’s fine for a few tools — but at scale, it means wasted tokens, slower responses, and higher compute costs.

Our proposal — Advertise-and-Activate — offers a simple alternative: let clients advertise tool summaries first, then activate full definitions only when needed.

This approach can:

• Cut token usage by up to 94% • Reduce latency and context bloat • Improve LLM reasoning and scalability • Require no protocol changes — it works today

It’s a small architectural shift with a big impact on performance and cost.

We’re already testing it internally across engineering workflows — from infrastructure management to developer tooling — and the results are promising.

This is a key piece of our broader vision for a scalable AI agent ecosystem. While our Agent Name Service (ANS) implementation focuses on secure, internet-scale discovery and verification, patterns like Advertise-and-Activate are crucial for making subsequent agent-to-agent communication via protocols like MCP both efficient and enterprise-ready.

Read the full article here: Smarter MCP Clients: A Proposal for Advertise-and-Activate: https://www.godaddy.com/resources/news/smarter-mcp-clients-a...

We’d love to hear your thoughts or experiences.

— Jay Gowdy & Travis Muhlestein

Building AI Agents from First Principles at GoDaddy

Everyone’s talking about AI agents lately, and for good reason. But at GoDaddy, we’re going deeper: starting from first principles to explore what makes an agent truly robust and usable in real-world scenarios.

Instead of asking “What can we build fast?” we’re asking “What design choices make agents flexible, testable, and reliable long term?”

Core Concepts

• Tool-centric design: everything an agent does is a tool call, with precise APIs and granularity. • Decision vs. delivery: agents decide what to do; tools handle how to do it—keeping systems modular. • Structured outputs & reflection: LLMs output both the tool call and the reason behind it, making debugging and iteration easier. • Universal tools: even user interactions (inform, confirm, request) are abstracted as tools, clarifying boundaries between logic and interface.

Real-world use cases → Not just theory

• Routing and responding to support messages • Surfacing emerging trends in sales data • Automating scheduling, inventory, or operations orchestration

What we learned

• Treating everything as a tool makes systems more predictable and extensible • LLM “verbosity” is valuable—it reveals reasoning and speeds iteration • Separating decision from execution reduces fragility and simplifies updates

We’re still at the beginning, but these principles give us a strong foundation. As agents evolve, architectural clarity matters more than chasing the latest framework.

Curious about architecture patterns that scale? Dive in here: Building AI Agents at GoDaddy: An Experiment in First Principles https://www.godaddy.com/resources/news/building-ai-agents-at...

Godaddy just introduced an implementation of the IETF Agent Name Service (ANS) to create trust through an open standard for agent identity and discovery that enables AI agents, businesses, and platforms to securely register, verify, and universally discover agents across search engines, AI assistants, and digital ecosystems. What are your thoughts?