TL;DR: We already take seriously the patterns for effective information flows for our humans, now we just need to apply this same thinking organisation-wide for our agents.

“I can’t trust the quality of what comes out of LLMs” and “I can’t trust the quality of what comes out of a new starter in my team” are equivalent concepts.

While in SF earlier this month, and from the 6 software companies I’ve worked with in July, I see the evolution of what we’re focusing on in the second half of 2026 for AI Engineering (and broader AI Knowledge work) as striving to solve LLM output quality at scale.

Organisations getting the best results (moving faster, maintaining quality) on this journey have two things in common: a Mindset Shift from 2025, and Approach to quality that’s (weirdly) unchanged from 2025, just applied to their agents.

  1. Mindset shift: The leadership and team members have all deeply adopted some variant of “I no longer do the task, I instead curate the context for my agents to do the task”.

  2. Approach: Instead of framing their task as “I need to write this document / create this plan / crunch this data / write this code”, they’re framing it for themselves as “When doing this task, I would historically have collected information from these people / codebases / slack channels / meeting notes, then I’d do it myself. Instead, I’m structuring that information for repeatable agent retrieval, and configuring tools to do the task to a measurable standard”.

What’s interesting to me here is basically every organisation already has this pattern: The moment a task is delegated to another person, we’re already immediately thinking “What information do they not have, that they’ll need to succeed?”.

Humans and LLMs are both probabilistic systems. Think about this scenario:

You have two people with 10 years of relevant professional experience. One of them has been working in your organisation for 3 years, the other is on their first day.

If you give them both the same high level brief for a task, and come back in 4 hours, you’ll get wild variance in the output quality. The person who’s been with you for 3 years has a wealth of explicit and implicit organisational knowledge, as well as the relationships within your org to know who to ask (and where to look) for the information they need.

The new starter with a similar professional history lacks this context, and there’s a good chance the quality will be “poor/unusable”.

At this point, we don’t say “I gave them this task, and they failed, so I guess I’ll keep doing it myself”. We instead fix the inputs by showing them where to find the missing information.

LLMs are trained on public information, so in this way, they’re like the new starter with 10 years of professional experience: Skills fit, but context miss.

Teach your teams to apply this “new starter” thinking to LLMs if they aren’t already. Then think about how cross-team communication currently works, and apply that to your agents too.