Agent, skill, or MCP? Which to use and when to use them | AWS’ Clare Liguori
Telling your agent “no” is a moat now, rearward deployed engineers, and harnessing the context for your SDLC
Every engineering team wants to build its own custom AI agent, but what if all your organization needs is a standardized skill or a stateless MCP server? This week on Dev Interrupted, Andrew sits down with AWS Senior Principal Engineer Clare Liguori to untangle the ecosystem of modern agentic architecture. They work through how a team decides which AI building blocks to own, and how to get that reach without inheriting a maintenance burden. Clare shares her perspective on the simplified MCP 7.28 spec and why stripping away heavy custom scaffolding is how enterprise AI scales.
That same shift is what makes MCP a gamechanger for LinearB customers, bringing your SDLC context layer, git, project management, and software delivery, into any agentic surface. What could your agents achieve if they can query your SDLC?
1. Preventing a proprietary tech monopoly
Tim O'Reilly gave some sage advice to the AI era this week, warning against relying too heavily on a handful of proprietary frontier labs. If everyone uses the exact same models with the same built-in personalities and guardrails, we risk stifling creative innovation and degrading to the lowest common denominator. The true key to an open source AI renaissance isn’t just open weights, but ensuring the modularity of the entire tech stack remains open. Initiatives like the AI Potluck are trying to build a rich, composable ecosystem for inference platforms and tooling so we don’t end up fracturing the community.
Read: Why Open Source Matters for AI
2. Giving models the keys to the machine
Anthropic is making auto mode the default permission setting in Claude Code. Up until now, this was an experimental feature you had to explicitly turn on. The system routes every tool call through a classifier to block irreversible or destructive actions on your local machine. Anthropic noted that users were manually approving 97% of prompts anyway, meaning developers were just reflexively clicking through warnings. While taking the human out of the loop speeds things up, we still need the ability to turn these safeguards off manually so our tools don’t refuse to act during critical security incidents.
Read: Auto mode is now the default in Claude Code for Pro, Max, and Team plans
3. Local agents heat up the desktop
Meta just dropped Muse Glimmer, a 30 billion parameter open weight model released under an Apache 2 license. What makes this long-running agent so compelling is that it only requires 20 GB of memory, meaning it can easily run on consumer grade local hardware without cloud dependencies. Through the power of distillation from a larger teacher model, it scored exceptionally well on benchmarks measuring the ability to synthesize long-form documents. The open source local model space is really starting to heat up, but remember that running an agent locally means you have to bring your own guardrails and harnesses to keep your environment safe.
Read: Introducing Muse Glimmer: An Open Agentic Model That Runs on Your Device
4. Engineering the ultimate context layer
There is a massive difference between a solo developer vibe coding a project and a team trying to build an enterprise product together. Agentic systems need immense amounts of relational context to make good decisions, and that context is usually scattered across your entire software development lifecycle. Without a clear map of that environment, agents are basically tourists. If you want them to act like native codebase contributors, you have to engineer a context layer that documents the current state, future plans, and risky components of your projects.
Read: Your SDLC is your context engineering
5. So you wanna build a software factory?
Are you ready to let agents triage, spec, implement, review and ship your code with humans entirely outside the loop? Before you automate your SDLC, you need to understand how to know it’s working.
On August 27th, I’m hosting a roundtable that brings Zach Lloyd (Warp), Dex Horthy (HumanLayer), and Dan Lines (LinearB) together to uncover what actually happens when you turn a fully automated factory on. We will discuss how to ensure your agents compound value instead of debt, and which parts of the PR lifecycle are actually worth automating first.
6. Spreading AI gains across the org chart
We have another life beyond tokenmaxxing story for you, but this one takes a completely different angle. Uber recently announced a strategy where they are sending their engineers out into other departments like legal, marketing, and sales. They're calling them "rearward deployed engineers," which definitely needs some workshopping as a title. However, the core idea is brilliant. If you have engineers delivering massive value with AI, enlisting them to act as internal consultants is the smartest way to distribute those productivity gains across the rest of the company.
Read: After starting the tokenmaxxing panic, Uber’s CTO is back with a very different AI story
7. The marginal benefits of custom developer setups
A new study exploring whether personalized skills actually help coding agents came to a rather surprising conclusion. After testing across hundreds of real-world developer sessions, researchers found that the productivity gains from highly customized, personalized skills were actually quite marginal. Instead, generalized skills that were pooled and distributed across many developers produced the largest and most consistent improvements. I've accrued a lot of personalized skills in my own workflows, but this data proves that if you want to scale impact, you need to consolidate what works and distribute those generalized learnings to the entire team.
Authored by: Shuyan Huang, Kai Du, and Andrew Lan
Read: Do personalized skills help coding agents? An empirical study of developer interaction histories
