Our own AI agent logs
Sort corrections by cause before blaming the AI agent.
AI work diagnostic
The tools you vibe-coded, finally working without you holding them together.
We find what breaks in your AI logs and build the fixes.
- Diagnostic
- USD 500
- With AI agent
- USD 1,000
Diagnostic in three business days of getting logs and access.
Sample diagnostic snapshot
Weekly updateBefore
You keep stepping in
- A person: Explain again
- A person: Fix output
- A person: Finish by hand
AI agent retries
After
You approve once
- An AI agent: Run skill
- An AI agent: Check result
- A person: Approve once
The team you would work with.
One team covers the technology, the management, and the people side of each workflow.
Our research so far: 11 research papers and 5 patents pending. This November we present at IEEE ICDM 2026.
Meet the full teamOur research
We study how people work with AI.
A production study, our own AI agent sessions, and a personal AI agent built from logs.
People
6,323
Using one production AI system in our study.
AI calls
197,858
Analyzed in the same study.
Working sessions
453
Sessions of three people with one AI agent. Our internal research.
Workflows
10
Packaged as skills from about six months of an entrepreneur's Claude Code logs, about 2 GB.
The study was accepted at the WMAI workshop held with IEEE ICDM 2026.
A personal AI agent we built
A member of our team turned an entrepreneur's logs into a personal AI agent reached by messenger. Repeating workflows became skills.
The study's de-identified dataset is public on Zenodo.
Your report is counted from your own logs.
The human side
An AI agent built around how you work.
Your tasks, your checks, your corrections: the starting point for your AI agent.
- A person: Sets the task — In your words, with the context you usually give
- An AI agent: Runs the skill — The workflow for this task, with its agreed check
- A person: Checks the result — Against what you agreed counts as done
- A named person decides: Accepts or corrects — Corrections you repeated in your logs are already built in
In our tests, we raise reliability up to twofold.
- How you set tasks. The context you keep retyping goes into the skill.
- How you check. Your acceptance criteria become a check: a test, a file, or a link.
- How you correct. A correction you made twice in your logs is already in the skill.
- What stays yours. You decide what the AI agent finishes alone and what needs your approval.
What the usual ways leave in place
- Writing the skills yourself. Memory misses corrections. The report counts them in your logs.
- Off-the-shelf skills. Generic instructions miss your tasks, style, and corrections.
- Keeping things as they are. The same task and the same correction, again next week.
The offer
The diagnostic alone, or with your AI agent.
Every figure in your report is marked: measured from your logs, estimated, or not available.
Diagnostic
USD 500
A report counted from your AI work logs.
For you if you want to build the fixes yourself.
- Repeat Task Map recurring tasks and how often they come up.
- Success Rate how often you accept the result without rework.
- Correction Map each correction sorted: a changed request, a clarification, a style preference, an input problem, or an AI agent fault.
- Done-Claim Check “done” claims with nothing to confirm them.
- Cost per Result time and tokens per accepted result.
- Fix List recommended changes, each with an estimated reliability gain.
Diagnostic and AI agent
USD 1,000
The report plus a working AI agent, ready for handover.
For you if you want us to build and check the fixes.
- The full diagnostic, included (USD 500 on its own).
- Skill Pack the repeating tasks we agree on in writing before the build, each packaged as a skill, plugin, or workflow.
- Done-Check Pack an agreed check per workflow. A test, a file, or a link backs “done.”
- Messenger Access your AI agent in Telegram or WhatsApp.
- Lives where you choose your own Claude Code or other tools by default, your cloud, or ours.
- 30-Day Fix Window if a skill breaks in the first 30 days after handover, we fix it free.
Examples of tasks that come back every week
- Email triage. Sort, draft replies, flag what needs you.
- Meeting notes to tasks. Decisions, owners, and dates into your tracker.
- Research briefs. A brief with sources for every claim.
- Weekly updates. Your week's work, ready for the team or investors.
- Follow-ups. Track replies owed. Draft the reminder.
- Code review. Check your rules. Record what was tested.
How it runs
From your logs to a working AI agent.
Agree the terms
Logs, dates, and checks, agreed in writing before anything is read.
Read the logs
We map recurring tasks and clarify corrections with you.
Report
Your figures and recommended fixes, each with an estimated gain.
Build and hand over
With the AI agent: checked skills in your chosen tools, connected to Telegram or WhatsApp.
Sending the form below commits you to nothing.
The terms
Fixed before work starts.
Price
Diagnostic: USD 500. Diagnostic and AI agent: USD 1,000. All AI usage for the work is on us.
Data
Separate written terms, agreed before work starts: logs, location, AI models, and deletion dates. Your records serve only the work you order. Any other use, including research or model training, needs your separate written permission.
Dates
The diagnostic within three business days of getting the logs and access. The AI agent build on dates agreed before the start.
Works-or-Free Guarantee
We agree a check for each workflow before we build it. If one fails, we keep working on it free until it passes, or refund you in full. Your choice. A changed request is new work.
Team
Run by the 8Hats Lab team above. Before work starts, we name the person who leads your diagnostic.
After handover
Runs on your own AI account in your own tools. For 30 days, if a skill breaks, we fix it free.
After handover, our cloud hosting and AI usage are billed separately, by agreement.
It fits when
- You give AI agents tasks every day, in Claude, Claude Code, Codex, Hermes, ChatGPT, or similar tools.
- The same tasks come back, and so do the same corrections.
- Your tools keep breaking. You step in to fix them.
- Your AI agent says “done” before the work is checked.
- Your sessions leave logs you can share under written terms.
Not a fit if
- You use AI a few times a week, mostly for one-off questions.
- No log of your AI work can be shared in any form.
For a process your whole team runs, see how we start with teams.
What if
Questions to ask before you start.
- “My logs are private.” Choose the logs covered. Exclude any project, chat, or period.
- “What if it does not work?” Free work until the agreed check passes, or a full refund. You choose.
- “It will cost more later.” Diagnostic and build prices are fixed. We check whether your logs support a diagnostic before starting.
- “I'll depend on you.” Your tools by default. Skills are written instructions you can read and change.
- “I'll be judged.” Corrections are sorted by cause. A changed request is not a mistake.
- “My setup already works.” The report shows it with figures. The diagnostic stands on its own.
Name your AI tools, your most frequent correction, and which offer you want. Leave out logs, files, and other people's personal data.
Start with the diagnostic
Contact us.
Tell us what you need. A member of the team reads every inquiry and contacts you personally.