Grounding, reviewing and trimming your way to something more reliable

The practice of reducing the number of tokens consumed by an agent's configuration and instruction files, so that more of the available token budget is preserved for the actual task, improving output quality and reducing cost. "You want it to be as token efficient as possible, aka have less text involved. So you wanna shrink down everything that we've just done and set up, get rid of everything you don't need that's in there. There'll be things that are left in there that you don't need. There'll be things that are wordy. Again, you can use AI to do this. You can take your configuration file, give it to an AI, and say, "I'm trying to make this more token efficient, but I still want all of these, all these configurations, measures that I put in place to be in place for this agent. How can I provide this in a more efficient way?""

In the context of AI, the context window is the context that an agent is pulling in and that is being sent to the model, including instructions, conversation history, and any provided documents or files.

Grounding is the practice of limiting where an AI agent can pull information from, constraining it to specific documents, URLs, or knowledge sources to control accuracy and reduce reliance on potentially outdated or biased external content. "Instead of just defining what is in your context window, you're defining what decisions your agent is gonna make to pull context in to your context window. And when you ground an agent, you think of that as limiting where your agent can pull information from. So if you ground an agent, for example, to only refer to these three URLs and don't refer to any other websites it's a good way to control the information that your agent is using and pulling into. This can be really helpful with things like guidelines that are regularly updated or there's multiple versions of, like WCAG guidelines."

How much does AI agent configuration matter?

An orchestrating AI agent responsible for receiving a large task, breaking it into subtasks, delegating those subtasks to specialised sub-agents, and keeping track of the overall progress, similar to how a delivery manager coordinates a team.Consider that different models have different costs. You can have your pilot agent on a very basic model. For example, you have a pilot agent whose job is to just make sure that each sub agent gets the right task at the right time, the first thing it can do is it can hand something off to a planning agent that that's on a better and more expensive model. The pilot agent gets all the information it needs ready, and gives it to the planning agent. That sort of prompt using a pilot agent is as token effective as possible.

Protecting sensitive data is probably the first thing you should think about whenever you're considering using AI for any purpose.Privacy by Design is an approach to building systems where protecting sensitive data is the primary consideration from the outset, rather than something added later. Security and risk mitigation are embedded into the design process, not bolted on.

A configuration file added to a local project (for example, .aiignore or .copilotignore) that tells an AI coding agent which files it cannot see or access, used to prevent exposure of secrets or sensitive data such as .env files.(NB. Caution is still paramount. This technique is not bulletproof as some AI coding tools can still see files explicitly stated in ignore files. For example, if the file is open in your IDE, or the file path is provided to the agent, or a tag such as @workspace is used.)

A technique where you provide an AI model with one or more concrete examples of the output you want, so it can base its response on those examples rather than relying solely on instructions.

How do we set up AI agents that are secure, reliable and trustworthy? Start with some principles.

A series of 99-second talks exploring careers, community, AI, accessibility, hiring, and the human side of quality engineering.