Agentic Work · RHAI
Your right hand doesn't just talk. It gets things done.
RHAI reads files, calls tools, and completes tasks on your behalf. Every action passes an approval gate, and every decision is written to a traceable ledger.
- Status
- Prototype
- Last verified
- 2026-07-18
- Access
- TODO_CONTENT_CONFIRMATION
The problem it solves
Daily work is full of repeatable steps — reading a file, gathering data, calling a tool, handing a result to someone. Delegating those steps to an agent saves time, but doing so without knowing exactly how much authority you've handed over carries real risk.
RHAI resolves that tension with one rule: the agent does the work, but every step that touches an outside system passes through your approval, and every decision is recorded in a way you can review later.
How it works
Type or say what you need.
RHAI runs low-risk steps — reading, summarizing — automatically.
At any step that touches an outside system, it stops and asks for your approval, right in the conversation.
The action runs; the decision and its reasoning are written to the ledger.
Verified capabilities
An approval gate, not a formality.
ActionGate denies any action that isn't defined, by default. A risky operation never runs without human approval, and each approval is a single-use, context-bound token that can't be stretched to cover another operation.
No surprises on the bill.
BudgetGuard caps spending before work begins — one ceiling per task, another per day. If a ceiling would be exceeded, the request stops before it's ever sent.
A traceable audit ledger.
Approval, denial, budget cutoff — every decision is written to a hash-chained ledger, so any later change becomes traceable rather than invisible.
Voice goes through the same gate.
A voice command or a scheduled task never holds looser authority than the text channel — the same rules, approvals, and budget checks apply across all of them.
It learns your work as you use it.
RHAI recognizes recurring workflows — read, summarize, send — and turns them into verified automations with your approval, without bypassing the gate or the ledger.
Architecture and integration
RHAI runs its agent runtime on a Go-based core and is reached through a web chat interface. As of today it runs in a production environment (rhai.uais.ai) and sees real usage — an internal verification recorded in July 2026.
Identity and access control, knowledge retrieval, and the data-access layer are provided by separate, specialized services; RHAI connects to them through defined contracts instead of embedding them.
Trust and control model
You always decide how much authority to hand over; RHAI starts with narrow permissions by default and denies anything undefined or out of scope without an explicit approval.
An approval is single-use and context-bound — it never stretches to cover another operation — and per-task and per-day budget ceilings are enforced before a request is ever sent, not after the fact.
Known limits / areas for improvement
SSE streaming can be degraded
Some real-time streams (SSE) can see reduced performance in the underlying model fabric; this is a known issue actively being worked on.
Occasional gateway 502s
The gateway can occasionally return a temporary 502 error; retrying the request resolves it.
Access model still settling
The product's formal access model and rollout plan haven't been confirmed by the product owner yet.
Evidence
Running in production
Verified on 2026-07-07: in the production cluster, both the core service (rhaid) and the web interface component were running, and the sign-in authentication flow was tested and confirmed working — running in a live production environment (rhai.uais.ai) with real usage, verified internally. This is an internal operational verification, not a published external report.
There is no published independent test report, benchmark, or customer case study yet. If you need concrete proof, the most direct path is a live technical call, where we can walk through the current setup and its limitations together.