TRUST LOOP FOR AGENTS THAT TRANSACT

Employ agents that evolve in your interest

Vijil helps enterprises employ teams of agents by giving them provable identities, measuring their alignment to the principal they serve, defending them against attack, and evolving them as their adversaries evolve.

Start for free See how Discover works

Trusted by

Recognized by

game of life?

Yes — but unlike the original Conway game of life, we model the behavior of two species, agents and adversaries, that compete against each other and co-evolve like a predator-prey population.

Both follow the same rule: a cell lives or dies by how crowded its neighborhood is. All we added is sides — a cell surrounded by the other species switches.

When one side gains more than 8% from parity, the trailing side gets reinforced. Whoever is behind runs harder, so falling behind is what provokes adaptation. Biologists call this the Red Queen: it takes all the running you can do to keep in the same place. We’re building agents that adapt to adversaries with the same zeal.

THE PROBLEM WITH SELF-IMPROVING AGENTS

Agents drift toward easier rewards.

An agent that transacts is scored on outcomes: deals closed, discounts won, orders filled. Improve it against those scores and it finds the cheapest route to them, including routes that run around or through your policy. No agent decides to break a rule. The population drifts, one reward at a time, toward whatever the scores pay for. The three roles that answer for the agent see it differently — and one person may hold all three roles.

Developer

The engineer who builds it asks:

“It scores higher every release, and I can’t tell whether it got better at the job or better at my tests. Give me a test it never trained on.”

Risk owner

The CISO who bounds it asks:

“An attacker needs one instruction my agent will follow. What stops a planted instruction from moving money or data — and can I show an auditor the control under the EU AI Act, NIST AI RMF and ISO 42001?”

Business owner

The GM who answers for it asks:

“The other side has an agent too, tuned to win. What can I safely delegate to mine, and what does one bad concession cost me? Show me the return with the downside priced in.”

Each intent needs its own control, and the controls have to share evidence.

THE TRUST LOOP

Vijil operates a trust loop around all the agents in your enterprise.

An agent that was safe last quarter is not safe now. The model changed, the tools changed, the adversaries changed. So the work of trusting it never finishes. Vijil runs that loop automatically. Discover gives every agent an identity it can prove. Diamond verifies it with tests on which it never trained. Dome holds the limits it cannot cross. Darwin evolves it inside them.

Each pass leaves evidence the next one uses: Diamond’s failures become Darwin’s targets, and Dome’s blocks become Diamond’s probes.

Use one module, or run the full loop.

Works wherever you build and run

Every Vijil module drops into frameworks and platforms you already use.

See the integrations

Frameworks

Identify

vijil discover

Every agent in your estate, with an identity it can prove.

·Scans code, cloud, endpoints

·SPIFFE identity per agent

·One inventory for all agents

Built on CNCF-standard SPIFFE

Verify

vijil diamond

How an agent behaves, before you trust it.

·Bespoke test suites

·Adversarial evals at scale

·Compliance-ready reports

Scored on tests it never trained on

Defend

vijil dome

What an agent may access, invoke and disclose.

·Labels follow data across every handoff

·Guards LLM and tool calls

·OTEL metrics built in

Mounts with two lines of code

Evolve

vijil darwin

A better agent, found inside your limits, for your approval.

·Searches variants within your constraints

·Hardens against production failures

·Proposes evidence-backed PRs

Every change re-verified by Diamond

Platforms

Try it now

Harden an agent in two lines

A git-style CLI: porcelain for the lifecycle, plumbing for control — and two lines of code to harden the agent from the inside.

install

$ pip install vijil-sdk# CLI + SDK

$ pip install vijil-dome# in-agent guards

python — harden from the inside

from vijil_dome import control

@control(policy="policy.yaml")# in + out

def call_model(prompt: str) -> str:

…

claude code

> /plugin marketplace add vijilAI/vijil# once

> /plugin install vijil@vijil# inside your session

Porcelainhigh-level lifecycle — one verb per step

discover → register → evaluate → protect → monitor → adapt → evolve ↻

vijil evaluate my-agent --baseline

vijil protect my-agent --guards prompt_injection,pii

Plumbinglow-level CRUD when you need control

agents · evaluations · harnesses · genomes · scores · policies · personas

vijil scores history my-agent

Dome, measured

Defends ten of eleven models, and the weakest the most.

On their worst 10% of responses, ten of the eleven score higher with Dome. The median change across all eleven is 17.6 points, and the weakest gain most: Mistral Medium 3.5 by 34.0 and Mistral Large 3 by 27.7. Kimi K3 does not move, 53.7 to 53.6.

Undefended

Domed

Twenty-two evaluation runs, one per model per state, each with its own run id. Undefended runs measured 3 and 4 September 2026 (Mistral Large 3 on the 8th), Domed runs 8 September.

Transform feral agents into fiduciary agents.

Start today. No sales call, no credit card. Dome is open source and needs no account: install it and guard a local agent in minutes. The hosted console takes a short form and opens the same day.

Start for free Have us run the first one

“Vijil helps us ship AI agents in six weeks instead of six months.” — Michal Nowak, SVP Engineering, SmartRecruiters

Vijil has raised $23M to build a platform that makes AI agents reliable, secure, and safe for enterprises.