Geordie AI

An agent stuck in a loop can burn through half a month's AI budget in a single day. Model routing that's a size too big, running at scale, can cost a business $1 million a month before anyone notices. The pattern's always the same: total spend is easy to see. Which agent, which workflow, which model choice actually caused it? That's the part nobody's tracking. That's basically the story behind Cost Intelligence, and why it keeps popping up in coverage - this time a shoutout from QA Financial in their roundup of AI and QA vendor news. More on how it works: https://lnkd.in/edNt22nS

Our Chief AI Officer Hanah-Marie Darley sat down with Ashish Rajan on the AI Security Podcast this week to discuss the reality of securing an agentic enterprise. Their conversation spanned a lot of big questions: - What does "AI-native" actually mean and why does it matter? - Why do agents break out of sandboxes? - How can you best understand the full agent lifecycle? Links to listen or watch in the comments

A skill and an MCP server can deliver essentially the same capability to an agent. What they cost to keep on hand is a different story. Take browser automation with Playwright. As a standing MCP server, its tool definitions can consume thousands of tokens in the agent's context before any work happens; as a skill, the relevant instructions can be loaded only when needed, potentially reducing that footprint dramatically. That difference may barely register on a single call, but standing context is carried across turns, so at scale it becomes a real line item. And the thing that lowers the bill - making a capability available only when it's needed - is also what can make an agent easier to audit. A leaner agent can be a cheaper agent and a more legible one. New piece works through the economics, with honest caveats about caching, for anyone who owns an agent budget or an agent risk register. Full read in comment below 👇

Employees burning tokens, no visibility into agent spend, no ability to connect that spend to business outcomes, and the first insight into cost is an invoice at the end of the month. It doesn't have to be like this. Geordie's Cost Intelligence gives AI strategy & ops, security teams, and finance the view they need to make smart decisions around agentic implementation

A pattern we keep seeing inside large enterprises: the 'golden catalog'. A curated, internally built set of pre-approved agent templates, signed off in-house so the business can deploy agents without depending on external agent builders. The instinct is sound. Owning your agent estate beats handing it to a vendor. The trouble is what it optimizes for. A catalog is a static asset, and agent capability is not static. The harnesses underneath it are shipping step changes in capability on a cycle measured in weeks. By the time a template clears the review and testing gates an enterprise rightly insists on, the model it was built against has often moved on. The catalog sits permanently behind, and the gap between what is approved and what is actually capable keeps widening. Owning your agent estate is worth doing. Owning a fixed set of approved templates is a snapshot of a moving target. Our latest blog dives deeper (link in comments)

Today, we’re launching Cost Intelligence - helping AI strategy, operations and finance teams connect agent spend to business outcomes. As agent pricing models change - and usage ramps up - tracking spend across different vendors and models has become increasingly chaotic. Cost Intelligence uses Geordie’s deep understanding of your agent architecture to give you a complete picture of spend in real time - and it traces that cost back to the productive work it powered. We’ve put a lot into this, so you don’t have to. Much more on our website: www.geordie.ai/cost-intel

Refusal training is one of the more reliable controls in the agentic stack: ask a model to do something harmful, and it declines. What it doesn't cover is anything that happens BEFORE the model gets a turn to reason. Claude Code's dynamic-context feature lets a skill run a shell command automatically at load time, while the harness is still assembling what the model is about to read. By the time the model sees anything, that command has already executed. This isn't hypothetical - Datadog Security Labs documented a skill, Clawsights, that used exactly this sequencing - the instructions could pass a human review and still carry the risk, because the part that mattered never went through a step either a human or the model was positioned to catch. Reviewing what a skill's instructions say isn't the same as knowing when it acts. Worth treating as a distinct question for anything already running in your environment, not just new skills going through review. More detail on the mechanism and what to check for in our latest blog, link in description!

Proud to be partnered with Amazon Web Services (AWS). Our VP of Partnerships Courtney Broadwell just got back from her second AWS Startup Partner Summit - a direct line to help shape how AWS supports businesses like ours. Geordie partners with AWS to bring AI agent security and assurance to our mutual customers – because the risk you can't see is the one that hurts you. #StartupPartnerSummit2026