HAI Human Agent Interface

Commercial HAI offers

Human Agent Interface Products

I turn messy AI work into scoped, controllable systems humans can actually own.

Scope the 4h HAI Setup Start small: Muenztelefon diagnostic Price on request after a short scope call

Tool-fit matrix

For which tools am I the right person?

I am not your general AI-coding-tool advisor. I am strongest when the problem is terminal-based agent work, control layers, project context, verification, and human ownership.

Area Tools / workflows Fit What it means
Terminal agents
Claude Code in the terminal, Codex CLI, Hermes
Very strong
Core competency: agentic work that lives close to the shell, repo, hooks, artifacts, and verification loop.
Terminal-adjacent agents
OpenCode, custom harnesses, human-in-the-loop workflows
Possible
Good if the problem is structure, scope, control, verification, or project context rather than product-specific support.
App-based agents
Claude Code Mac App, Codex App, chat-style agent UIs
Limited
Methodically useful, but weaker for deep tool setup when the workflow is mainly inside an app surface.
IDE-based AI tools
Cursor, Lovable, Google Antigravity, vibe-coding platforms
Not focus
Usually not the right fit for deep product advice or tool support. I can still help with ownership, gates, and workflow boundaries.

HAI does not mean knowing every AI-coding tool perfectly. HAI means structuring agentic work so the human keeps orientation, control, and ownership.

Not sure? Scope your setup

Why the terminal-agent row is marked very strong: these are local evidence snapshots from real work. They combine the Claude Code backup before the April reset with the current post-reset usage data, then deduplicate by session id.

Codex HAI depth 38 repo sessions
2,627 tool calls

Strict Human-Agent-Interface repo sessions from 2026-05-11 to 2026-05-14, with 163 turn records and 7,476 response items.

Hermes insights 369 sessions
282.6M tokens

Last 30 days: 3,928 tool calls, 39 distinct skills loaded, and a manual CLI slice of 61 sessions with 7,018 messages. See Hermes system.

Eval discipline 166 runs
44 / 44 gate eval

AgentArena has 166 runs, 9 tasks, and 37 agents. Sidecar-NG adds 37 prompts and 97.3% internal local-eval accuracy.

Four real offers

Pick the product that matches the shape of your agentic problem.

The core path is still agent-native. If the problem is earlier than that, start with a small diagnostic or orientation call before a full audit is priced.

Offer 01 / Paid setup

Personal HAI Audit

A 4h teardown and rebuild of how you currently work with agents.

Who it is for

Founders, builders, developers, and operators with terminal-based or terminal-adjacent agent workflows, especially Claude Code, Codex CLI, Hermes, OpenCode, or custom harnesses.

Problem it solves

You have momentum, but the work fragments into chats, half-plans, unclear decisions, and agent output you do not fully own.

What happens

We inspect the current workflow, identify failure points, separate human decisions from delegable work, and redesign the next agent loop.

Deliverables

  • Workflow diagnosis
  • Rewritten next-action packet
  • Tool/setup changes to make immediately
  • Risks, limits, and stop rules
Price after short scope Scope this setup
Offer 02 / Build the system

Agentic System Setup

A practical setup for people who need a working harness, not another AI strategy note.

Who it is for

Technical founders, small teams, and serious individual builders who want repeatable agentic work across projects.

Problem it solves

Your tools can run, but the operating system around them is missing: roles, permissions, artifacts, verification, and handoff.

What happens

We design the minimum agentic operating loop, configure role boundaries, and define how work moves from request to execution to verification.

Deliverables

  • Agent role map
  • Bounded execution workflow
  • Verifier path and definition of done
  • Handoff packet for repeated use
Best after one audit or clear system pain Scope this setup
Offer 03 / Prototype-backed layer

Sidecar / Sidecar-NG

A watcher and downscope layer for long, risky, or overloaded human-agent sessions.

Who it is for

Heavy agent users who lose focus, over-delegate, miss stop points, or need a second system watching the collaboration itself.

Problem it solves

The main agent optimizes for doing work. Sidecar watches the interaction: scope, overload, drift, verification gaps, and human decision points.

What happens

We map where your sessions fail and decide whether a Sidecar-style watcher, prompt layer, or workflow rule set is the right intervention.

Deliverables

  • Sidecar-fit assessment
  • Watcher rules or prompt layer
  • Downscope triggers and intervention points
  • Prototype path if a custom layer is worth building
Prototype direction, not mature SaaS Read Sidecar proof Check Sidecar fit
Offer 04 / Team architecture

Adminspace / Userspace

Separate productive agent use from governance, review, permissions, and system evolution.

Who it is for

Companies and technical teams where multiple people or agents touch the same workflows, repositories, or operational decisions.

Problem it solves

Teams blur user work, admin work, security decisions, prompt changes, review, and tool access until no one owns the risk.

What happens

We draw the boundary between userspace and adminspace, then define who can change agents, approve work, inspect evidence, and stop runs.

Deliverables

  • Userspace/adminspace separation map
  • Permission and review model
  • Governance and escalation gates
  • Next pilot slice for a real team workflow
Company/team discovery first Map team system

First commercial step

Start with scope. Not a price wall.

You bring one real agentic workflow problem, not a hypothetical transformation program.

We leave with a concrete setup, packet, boundary model, or decision about what not to build.

If you are not agent-native yet, start smaller with orientation or a Muenztelefon diagnostic.

The price comes after fit is clear. Early feedback pilots can be scoped differently.

FAQ

Commercially clear, technically honest.

HAI is early, but not vague. The first sale is a setup session. The deeper products grow from real workflow evidence.

Is this consulting or software?

Today, the first offer is a paid setup and audit. Some parts are software-backed, especially Sidecar-NG and evaluation work, but the sale starts with the human workflow.

Who is this for?

The core fit is people already working with agents who feel the cost of unclear scope, context drift, weak verification, too many threads, or tool setups that do not fit how they think.

Can I start if I am not agent-native yet?

Yes, but the entry stays narrow: orientation, trust, boundaries, and one diagnostic next step. HAI does not become a generic AI beginner course.

What do I get after 4 hours?

A diagnosis of the current setup, concrete workflow changes, a next-action packet, and a clearer boundary between human decisions and agent-executable work.

Is Sidecar-NG a product or a prototype?

It is prototype-backed product direction. The benchmark result is internal evidence, not an audited market claim. The setup decides whether a watcher layer fits your case.

Can this work for a company/team?

Yes, but the team version starts by mapping userspace, adminspace, ownership, review, permissions, and stop rules before any larger rollout.