Using AI agents makes me always rebuild the same prompts. Session after session, the same instructions: Do not write verbose comments, review after implementation finishes, use these conventions, dos and donts. And the set is changing every time a new model tier is released. The workflow still needs to be designed every time.
The research has a name for this, and it is backed by real data.
What the research says
The Microsoft Research and Carnegie Mellon University study published at CHI 2025 (source) surveyed 319 knowledge workers on how generative AI affects their critical thinking. The headline finding was that 40% of tasks involved no critical thinking applied to AI output - users accepted results without scrutiny. But a more subtle finding matters more for this conversation: the cognitive work shifted from generation to evaluation. Instead of figuring out how to solve a problem, users described what they wanted, then evaluated what the AI produced. Both modes are draining, but they drain different cognitive resources.
The Harvard Business Review study published March 2026 (Bedard, Kropp, Hsu, Karaman, Hawes, Kellerman - source) went further. Surveying 1,488 full-time US workers, they found 14% of AI-using participants reported "brain fry" - a state of mental exhaustion tied to specific usage patterns. The pattern that correlated most with burnout was task-reinitialization: re-explaining context, re-specifying constraints, re-defining the task boundaries every time you sit down to work with an AI tool.
A 2025 study in Frontiers in Psychology (Chirayath et al., source) described the mechanism: AI tools enable cognitive offloading (using external aids to reduce mental effort) but can also create cognitive overload - an erosion of introspection, over-reliance on algorithmic feedback, and anxiety induced by the need to constantly specify and re-specify context.
These studies converge on one pattern which is the overhead of re-initializing an AI agent session is not just annoying - it is measurably fatiguing in ways that compound over the workday. Each how-to-correct prompt scaffolding costs a small cognitive tax. After the fifth or sixth session, the tax dominates the value the agent delivers.
Why the fatigue compounds
The model gets smarter every release. But smart models still need well-designed workflows to produce consistent end-to-end results. A model knows code review patterns in theory. It still needs to be told which patterns your project uses, what conventions to follow, what to ignore, and how to format the output.
Partial solution for this is Skills.
A skill encodes that workflow design once. The skill says: this is the trigger (a PR is opened), this is the methodology (check for these 7 things in this order), this is the output format (inline comments with severity labels). The agent discovers the skill when the trigger fires. The workflow is executed without the user re-specifying the constraints.
But writing a skill for a complex workflow which works from the start is almost impossible. It needs tuning, validation. Once new model releases with its own behavioral quirks - your skill becomes outdated.
How Agent Armory solves this problem
When you connect your harness to Agent Armory, your agent now has access to all the skills in the library. You can stop thinking about writing a skill yourself or searching for the best skill online. You just write to your agent "Do this task with Agent Armory".
Your agent will find and use the best applicable skill and submit the deliverable for your review. If you are satisfied - fine. Your harness will send a positive signal. If you are not - Agent will send issues it faced and proposed improvements. Next time you use the skill it is already updated. And if you dont use the skill for a month, others will - and skill is fresh again.
Our catalog has more than 4000 skills. Both open licensed sourced from web and our own: tuned and grounded, verified on real work. Every one of them is a prompt you would otherwise have written from scratch.
AgentArmory
Register, copy mcp URL and add as connector to your harness and you are good to go.
If you don't want to prompt "with AA" yourself, use AgentArmoryinit skill to modify your Claude/Agents/Soul md file.
That is the entire setup. One MCP endpoint, tuned and fresh skills. The agent discovers them by capability at runtime. See the full list on the website.
How it maps to the fatigue research
The Microsoft study found that the shift from generation to evaluation was the cognitive drainer. Skills move the work back to evaluation: the skill defines the generation approach, and the user evaluates whether the output meets the standard. But the initial generation - the "how do I approach this" phase - is no longer rewritten from scratch each session. The skill already encodes the approach. You are evaluating a structured output against a known standard, not inventing the approach from scratch.
The HBR "brain fry" study identified task-reinitialization as the highest-correlated burnout pattern. Skills eliminate reinitialization entirely. The skill is installed once. Every session after that, the agent already knows how to handle the task class. The cognitive load of setup is zero.
The Frontiers study warned about cognitive overload from needing to constantly specify and re-specify context. Skills bound the context: the skill says what it does, the user confirms intent, and the agent executes within the skill's defined scope. No drift into undefined behavior, no re-specifying boundaries because they were already specified in the skill definition.
Where this is heading
The model gets better. The workflow should too. Here is the direction the library is moving: run a skill, comment on what went wrong, your agent suggests an improvement - and so do a hundred other people's agents using the same skill. The skill refines itself across users, not just across your sessions. Next time you install it, the edge cases a dozen other teams already hit are handled.
(Our own engineering team already follows this pattern: when a QA pass catches an issue in one of our first-party SWE skills - Implementation Workflow, Task Decomposition, Code Review Workflow - the fix gets folded into the skill definition before the next install. The public-catalog feedback loop that lets any user's agents suggest improvements to any skill is in development.)
Try it tonight
Pick one task you wrote more than three explicit prompts for this week. Connect Agent Armory. Use "Solve the task with Agent Armory" instead of writing the prompt again. One session is enough to feel the difference.
Top comments (0)