In short: The Ecuhawk agent is an AI teammate built on the Ecuhawk engine and the open-source Hermes agent. It controls the bench through Ecuhawk's own tools, works on its own on a schedule, asks a person before any action on real hardware, learns from every run, and talks to your team where they already are. It is part of Ecuhawk at launch and available to beta teams now.
Bench control, test writing, diagnosis, approvals, fact checks and the audit log run today in the private beta. Chat apps, team tools, schedules and learning are part of the agent at launch and are switched on for each team during setup.
What the agent can do
Run the bench
Checks bench status, powers the ECU, runs single tests or whole suites, reads signals and gives verdicts with measured values.
Write tests from requirements
Reads your requirement documents and writes test cases in Ecuhawk's own format, checked automatically and saved as drafts for review.
Diagnose the ECU
Reads identity and fault codes over UDS, and searches the trace for missing frames, timing problems and odd values.
Ask before acting
On real hardware every write needs a person's approval. Hard limits always apply, and the emergency stop never waits.
Check its own answers
Every number and name in a report must match a value the tools really returned, or it is flagged before anyone sees it.
Keep a full record
Every tool call, approval and refused action is written to an audit log, ready for reviews and quality audits.
Work on its own
Runs jobs on a schedule: nightly regression, morning bench check, weekly coverage summary. Results go out without anyone asking.
Learn your bench
Keeps memory across sessions and turns solved problems into reusable skills, so it gets faster and more accurate on your ECUs over time.
Talk where your team talks
Telegram, WhatsApp, Slack, Microsoft Teams, Discord, Signal and email. Ask for a run from your phone and approve it with one tap.
Work with your tools
Jira and Confluence, GitLab, Linear, Notion, Grafana, Sentry, CI servers and more: draft bugs, update pages, attach reports.
Split big jobs
Hands parts of a large job to helper agents that work in parallel, such as one per bench or one per requirement group.
Use your model
Cloud models through your own account, or open models on your own server so nothing leaves your network.
How it works
1 · Ask or schedule
A chat message or email, a Jira ticket, or a timer such as "every night at 2:00".
2 · Plan
The agent picks the tests, signals and steps from your project files and what it learned before.
3 · Safety gate
Simulated bench: it acts. Real bench: it asks a person. Outside the limits: refused, always.
4 · Report
Verdict, measured values, plots and a plain-language explanation, sent to the right people.
Built to be trusted on a real bench
| Bench mode | Reading | Writing and running |
|---|---|---|
| Simulated | Yes | Allowed, within bench limits |
| Real hardware | Yes | Only after a person approves. Diagnostic services other than reads are blocked. |
| Offline (log files) | Yes | Refused |
- Drafts, not surprises: AI-written tests are saved as drafts, never overwrite reviewed tests and run only on a simulated bench until a person accepts them.
- Hard limits in the engine: a request for 20 V on a 16 V bench is refused by Ecuhawk itself, whoever asks.
- Allow-lists: reports and messages go only to people and channels you approve.
- Your data stays yours: see AI and your data.
Proof: real runs on our bench
Real results from the agent on our simulated bench in October 2026, using a Google Gemini model.
"Check the bench and run all checks"
2 tool calls, 23 seconds, all checks PASS, report saved.
"Read the ECU identity, then try 20 V"
Found the ECU unpowered and switched it on, read the VIN and software version, and the engine refused 20 V as outside the 0 to 16 V limit. Sleep-current check PASS with 7.8 % margin.
"Read the requirements and write the tests"
6 valid test files with 0 problems. 5 passed at once; one used a wrong expected voltage and was corrected in review.
We also show the mistakes. In one run the agent added data it had never read. That is why every answer now goes through a fact check, and why AI-written tests stay drafts until a person accepts them.
Questions
Can I control Ecuhawk from Telegram or WhatsApp?
Yes. The Ecuhawk agent can be connected to Telegram, WhatsApp, Slack, Microsoft Teams, Discord, Signal or email during setup. You can ask for runs and approve actions from your phone.
Can the agent create Jira bugs?
Yes. With Jira connected during setup, the agent drafts a bug for a failed test with the report and trace details attached, for a person to review.
Does the agent learn over time?
Yes. It keeps memory across sessions and saves solved problems as reusable skills, so it becomes faster and more accurate on your bench.
Can it run tests without anyone asking?
Yes. Scheduled jobs such as a nightly regression or a morning bench check run on their own and send results to the people you choose. Actions on real hardware still follow your approval rules.
Will the AI change things on real hardware without permission?
No. On a real bench every write needs a person's approval, the engine refuses anything outside the bench limits, and the emergency stop always works.
Which AI models does it use?
Common cloud models through your own account, or open models on your own server. Our own runs used a Google Gemini model.