Rules, not guesses
Every action an agent takes is checked by a deterministic rulebook before it runs. Same input, same decision, and the rule that decided is on the record. A model can be talked out of its instructions; a rule cannot.
About Abhed
We make AI agents answerable.
AI agents now refund customers, change records and send data on their own. When one gets it wrong, someone has to be able to say who allowed it, why, and prove it. Abhed is that someone’s system.
The name
Abhed means impenetrable.
Not a wall that stops work, a gate that nothing passes without a reason. Most actions go straight through in milliseconds. The few that matter wait for a person. Every one of them is written down where no one can quietly change it.
Why it exists
Agents got permissions. Accountability didn’t come with them.
Companies give AI agents the same access a person has: refund up to a limit, email a customer, export a report. But a person can be asked why. An agent’s mistake usually surfaces days later, in a complaint or an audit, with no clear owner and logs anyone could have edited.
The usual answers are all-or-nothing: approve every action (and the agent becomes useless) or approve none (and one bad action costs lakhs). Text filters on the model’s output don’t catch a wrong number or a hidden instruction in a customer’s message.
Abhed sits where the agent’s decision turns into an action. It checks the action, not the words, and keeps people in charge of exactly the moments that need them.
How we think
A person for the big moves
Small, reversible work runs on its own. Refunds above a limit, data leaving the company or a change to access wait for the right person, who approves on their phone with the evidence in front of them.
Proof that shows any edit
Each decision is a record chained to the one before it, sealed in signed blocks and timestamped outside the company. If anyone changes the past, the exact record shows it.
Right now
This isn’t a mock-up.
These numbers and decisions come from the running system, refreshed every few seconds.
—companies connected
—agents governed
—decisions in the last 5 min
—gate decision, p95
Latest decisions
Connecting to the live ledger…
What we built
The gateway
Agents call it before they act; it decides in milliseconds.
The rulebook
Plain-English rules per company, versioned, two people to change one.
The approver app
A phone app for the people who sign the big moves.
The ledger
Hash-chained, signed and anchored, verifiable by anyone.
No-code connect
A company points its agents at Abhed with a base URL and a key.
Break Abhed
Anyone can try to make our agents misbehave, live.
The team
Five people, one gate.
Abhed was built by a team of five for CodeX 3.0. Security and the backend, the screens, the research behind the design, and the business case each had an owner.
Harsh Dubey
AI security, guardrails & backend
Owns the part of Abhed that says yes or no. Every action an agent wants to take passes through the gate he built, and the decision is made by rules, not by another model that could be talked out of them.
What they did
- The deterministic rulebook and the gateway that checks every tool call in about a millisecond
- Guardrails against prompt injection, bulk data export, destructive and privilege-changing actions
- The no-code AI gateway: companies connect with only a base URL and a key
- The backend: accounts, sessions, the hash-chained ledger and the APIs behind every screen
- AI security
- Guardrails
- Gateway & connections
- Backend
Naitik Gupta
Frontend, testing & deployment
Makes Abhed something a person can actually use, and makes sure it keeps working. Built the screens judges and companies see, tested each flow end to end, and put the whole system live.
What they did
- The fleet monitor, command center, connect page and the approver app people use on their phones
- Light and dark themes, phone layouts and the checks that keep text from overlapping on any screen
- End-to-end tests of the demo flows: a held refund, an approval on the phone, a blocked export
- Deployment of the app and the three demo company websites
- Frontend
- UI
- Testing
- Deployment
Naman Verma
Backend & presentation
Works on how data moves between the agents, the gate and the people who approve, and explains the system so that anyone can follow it in a few minutes.
What they did
- Backend services and APIs for agents, approvals and the company websites
- The data flow from an agent’s request to a recorded decision
- The presentation deck: the problem, how Abhed works and the live demo story
- Backend
- APIs
- Pitch deck
Riya Mishra
Research & architecture
Started from the question of where AI agents actually go wrong in companies, and designed the architecture so the gate sits exactly where an agent’s decision turns into an action.
What they did
- Research on real failure modes of AI agents: wrong amounts, hidden instructions, data leaving the company
- What accountability regulators and auditors expect, and how the ledger proves it
- The system architecture: gateway, rulebook, human approval and tamper-evident records
- Research
- System architecture
- Compliance
Manshi Chaudhary
Market, go-to-market & pricing
Answers who needs Abhed, why now and what they would pay. Turned the product into a business case that a bank, a hospital or an online store can say yes to.
What they did
- Market gap analysis: what existing AI governance tools miss and who is left without one
- The go-to-market plan: first customers, sectors and how a pilot starts
- The pricing model: per agent each month, by risk level, so the price grows with the value
- Market gap analysis
- Go-to-market
- Pricing
Watch an agent ask for permission.
Open the live demo, or connect your own agents with a base URL and a key.