Yossi Yehezkel
CEO
- Designed and built YMUX.
- Founder of WeDeliver il (20+ paying customers).
- n8n Official Ambassador in Israel.
AI coding agents that banks can audit.
AI writes the code. YMUX controls what gets merged.
PersonAgentCode
Audit recordREQ-0142“Add CSV export to the reports page.”
- RequestTicket opened in plain words
- Goal3 outcomes. Confirmed by a person.
- Plan + testsPlan and tests approved by a person
- Build + checkBuild, lint, types, tests: pass. Review agent: no block.
- Sign-offWaiting for a person
Merge lockedMergedNo AI model can merge. Waiting for a person to sign off.A person signed off. YMUX merged and saved this record.
Why AI stays out of production code
- Regulated companies need an approver and a test record for every code change.
- Today’s AI coding tools help one developer write faster. They do not give the company a fixed process with approvals and proof.
- So AI stays out of production code.
How it works
Five steps for every change. Each step shows who acts: a person, an agent or code.
Step 1: Request
An engineer opens a ticket in plain words. Example:
Add CSV export to the reports page.
Step 2: Goal
An agent restates the request as a goal with outcomes. A person confirms it. If the agent misread it, we log why: missing context or unclear wording.
Step 3: Plan + tests
An agent writes a plan and the tests that prove “done”. A person approves them.
Step 4: Build + check
Agents write code on a separate branch. Build, lint, types and tests run after each step. A second agent checks the code against the plan.
Step 5: Sign-off
A person signs off. YMUX merges and saves the record.
What is deterministic
The workflow engine is ordinary code, not an AI model. It chooses the next step, retries, stops to ask a person, and merges.
No AI model can merge code on its own.
Next stepWorkflow engine (code)
RetryWorkflow engine (code)
Stop and askWorkflow engine, then a person
MergeWorkflow engine, after a person signs off
Write codeAgents, inside their file limits
What you get for every change
One audit record per request. This is what your auditor opens.
Audit recordREQ-0142Sample record. Values are illustrative.
Request and goal
- request
- “Add CSV export to the reports page.”
- goal
- Reports page has an Export CSV button. The file contains the rows in the current filter.
- misread
- none logged
Approvals
| Step | Approved by | When |
|---|---|---|
| Goal | Tech lead | 09:20 |
| Plan + tests | Tech lead | 09:41 |
| Sign-off | Eng. manager | 11:02 |
Checks
| Check | Result |
|---|---|
| build, lint, types | pass |
| tests for this ticket | pass |
| checks from past tickets | pass |
| review agent vs. plan | no block |
Model and cost
- model
- Claude, via Amazon Bedrock
- tokens
- 1.9M in, 84K out
- cost
- $4.72, all retries included
Files and limits
- may change
- src/reports/**, tests/reports/**
- changed
- src/reports/export.ts
src/reports/ReportsPage.tsx
tests/reports/export.test.ts - outside limits
- 0 files
What makes YMUX different
The process is code.
The same request always follows the same steps and checks.
Agents are not trusted.
Work moves forward only when automatic checks pass. The review agent can stop a merge. It cannot approve one.
Each agent has limits.
Each role has a written list of files it may read and change, enforced during the run.
Old features stay proven.
Checks from past tickets run again before each new merge.
Less review, same control.
Optional. When a step is right 90%+ of the time, the customer can switch it to one final approval. Default: a person approves each step.
Why not GitHub or Cursor?
GitHub controls the merge button. YMUX controls the path before it.
Goalapproved
Plan + testsapproved before code
Runwrite limits per agent
Past checksrun again
Recordone per request
Merge buttonbranch protection, required reviews
YMUX
GitHub
Cursor and similar tools help one developer write faster. YMUX gives the company the fixed process around them.
Works with any agent. Runs on your servers.
Proof so far
Current focus: bug fixes, new features and research tickets.
- 280
- tickets completed in 30 days on our own 2 products
- 4.1 h
- median from request to merged code (agents work ~1 h of it)
- ~$5
- AI model cost per completed ticket, all retries included
- 47
- tickets completed in one 28-hour run
Our own production use, Sep 8 – Oct 7 2026. No external customers yet.
Team
Dr. Yaniv Bouhadana
CTO
- PhD, Bar-Ilan University.
- 12+ years of full-stack work.
- Former Tech Lead at Citi.
- Builds multi-agent AI systems.
3 of 3 seats open
Looking for 3 design partners.
Regulated engineering teams that want AI in production, with proof.