Labs · Models, agents and product
Improve agents through real work.
I review specific features and human–agent workflows. With your team, I turn usability problems into inspectable findings, design requirements and proposals for fixes.
This page is for teams building AI agents. For a small business, see consulting.
- Start from real work, not a benchmark
- A targeted review, quoted after a first conversation
- Public sources for every contribution
Method
From a situation to an inspectable fix.
Three steps, from the work encountered to a reproduction the team can read back.
- 1
Observe Start from work to be done and the difficulty encountered, without publishing private exchanges. Describe the situation
- 2
Diagnose Compare intent, actions, observable state and what the interface communicates. Read the analyses
- 3
Reduce and verify Produce a reproduction, requirement or fix; state what was tested and by whom. See a contribution
case/0012
Private exchanges are not published. Public cards cite their sources.
Reduced cases
case/0012
claim_supported: false
Hermes Bot Screen
5 findings reproduced
Private exchanges are not published.
Labs formats
A targeted review is quoted after a first conversation, from €1,500 excl. tax per day. The detail is in the form.
One feature or human–agent workflow, examined with the team.
| Format | Deliverable | Scope | Price |
Targeted review QuotedOne feature or human–agent workflow. Quoted from €1,500 excl. tax per day, after a first conversation. | One workflow | Quoted |
eval-case v1
A reduced failure case: claim, action, evidence
Documentary example, no replay verdict
- Status
- No replay verdict
- Publication
- Published
| Case | Status | Family | Publication |
Documentary example | No replay verdict | FCE | Published |
Code contributions
Reviews and fixes already merged, with sources
| Contribution | Status |
| Hermes Bot Screen | Merged |
| Treg | Merged |
| Hermes × xAI | Merged |
Starting points
When should you request a review?
An agent claims a result without evidence, its work is hard to take over, or an integration blocks the intended use. The examples below show what a review can examine.
Claim
“Done.” But nothing proves it.
Distinguish a claim from a verified state. case/0012 is a documentary example, with no replay verdict.
Control
Someone must monitor and reconstruct the work state.
Make human control and takeover clear. Five documented issues, reproduced by the maintainer, feature merged on 23 September.
Integration
The integration blocks the intended use.
Forward settings and respect access scope. The Hermes × xAI fixes are one contribution chain, not a partnership.
Further reading.
When “Take over” Doesn’t Give You Control
Read the article
Why I’m Contributing to AI Agents as an Ergonomist
Read the article
Raw Traces Are Not Evals
Read the article
The user’s head is not the runtime
X · English
Grok Bot
I · X · English · II · X · English
Real delegation · 24 August 2026
X · English
Hermes Bot Screen
Design and reliability review
Treg
Code contribution, reviewed and merged
Hermes × xAI
Code carried forward with attribution · April–May 2026
Open source
Documented contributions.
From control handover in Hermes to read isolation in Treg: identified problems, reviewed contributions and integrated changes. I have worked with the xAI teams and the Nous Research teams.
Hermes Bot Screen
Who controls the browser? Human takeover must suspend agent actions, and the screen must make its state clear. At Teknium’s invitation, I reviewed the human–agent interaction. I documented five issues: actions during human control, stale capture across a takeover cycle, control released on unexpected disconnection, old images presented as live, and profile collisions. Two design notes complemented the review. The maintainer confirmed reproducing all five issues, documented fixes and published verification results. The feature was merged on 23 September. The cited verification is the maintainer’s, not a user study. The design notes are not two additional shipped features. The handoff-reason mechanism was later removed. Takeover does not wait for in-flight actions to finish. PR #108914 · Review · Julien · Reproduction & fixes · In-flight actions · Handoff revision
Treg
A client identity pinned to a connection must also limit what it can read back: history, results and asynchronous tasks. I proposed a fix applying that identity to the relevant read paths and separating replay namespaces. The maintainer reviewed the PR and completed missing paths. PR #616 was merged after two maintainer follow-ups, including routed calls, refusals and associated feedback. The scope is read isolation under the pinned identity. This establishes neither general isolation of all accounts and public media nor deployment of this version. PR #616 · Julien · Review & follow-up 1 · Review & follow-up 2
Hermes × xAI
Using xAI in Hermes requires native integration and correct forwarding of the selected reasoning effort. My native-provider and reasoning-effort contributions were carried forward in follow-up PRs, with attribution preserved. Native integration was merged on 10 April in #7372; reasoning-effort forwarding on 9 May in #22807. Original and follow-up PRs form the same contribution chain. These fixes do not establish a general performance gain, a partnership with xAI or use of my work to train its models. Native · #7372 · Attribution · #7050 · Reasoning · #22807 · Attribution · #22055
Propose a feature for review
Preparing a feature, revisiting a workflow or investigating a usability problem? Describe the product, the workflow and the decision you need to make. We can scope a focused review: findings, reproductions where useful, recommendations and a discussion with your team.
First conversation: 30 minutes, by video or by phone, free and without commitment. Reply within 48 business hours.
Choose how to get started
Propose a review
A conversation to scope a targeted review.
- The product and the workflow
- The decision to inform
- A quote after that first conversation
Read the public cards
The contribution cards, distinct from a client relationship.
Teams I have worked with
SpaceXAI and Nous Research. Each logo leads to the case study.
Frequently asked questions
case/0012