Platform/Darwin
VIJIL DARWIN · EVOLUTION

Evolve self-organizing agents

Trust decays as the world changes. Darwin adapts your agents to attacks and failures observed in production — and every change arrives as a pull request your team approves.

Start for free Learn more
$ vijil evolve --spec spec.yaml
Darwin execution traces
DARWIN · PROPOSAL
PR #142 — harden my-agent
trust score82 → 91
re-verified✓ DIAMOND
awaiting reviewILLUSTRATIVE
TRUST DECAYS

An agent that was safe last quarter is not safe now

The model changed, the tools changed, the adversaries changed. A static agent does not hold its score; it loses ground to a world that keeps moving. Re-testing tells you how far it has slipped — it does not close the gap.

The ground moves

New model version, new tool, new attack technique. None of them wait for your next release.

Findings rot in a backlog

An evaluation that produces a report produces work — work nobody has time to do.

Fixes do not compound

The same class of failure gets patched by three people in three places, and the fourth one ships.

HOW IT WORKS

Failures become pull requests

PRODUCTION attacks · drift · failures Observe telemetry in Analyze root cause Mutate targeted edits PULL REQUEST + hardened prompt − permissive tool scope nothing ships until your team merges ↻ re-verified by Diamond before merge PRODUCTION attacks · drift · failures Observe telemetry in Analyze root cause Mutate targeted edits PULL REQUEST + hardened prompt − permissive tool scope nothing ships until your team merges ↻ re-verified by Diamond before merge
Darwin verifies every mutation and requires developer approval to merge — continuous improvement, not autonomous self-modification.
Read the Darwin docs →
WHAT DARWIN MUTATES

Three strands of the agent genome 

CODE
Agent logic, prompts, and tool-use flow.
CONFIGURATION
Model choice, parameters, tools and permissions.
CONTROLS
Guardrails and policy thresholds.
Darwin architecture notes →
INSTALL & RUN

It opens the pull request; you merge it

Darwin reads what Diamond found and what Dome saw in production, proposes a change to the agent’s prompt, config or code, and opens a pull request against your repo. Nothing lands without a human approving the diff.

evolve
$ pip install vijil-sdk# CLI + SDK
$ vijil evolve --spec spec.yaml# propose mutations
$ vijil adapt my-agent# learn from production
PULL REQUEST, NOT A PATCH
review the diff like any other change
THREE STRANDS
prompt, config and code — the agent genome
TWO SIGNALS IN
what Diamond found, and what Dome saw in production
PROOF

Darwin keeps trust from sliding

100 70 40 0 weeks in production minimum trust to operate without continuous improvement with Darwin new attack model update new rule data drift ILLUSTRATIVE 100 70 40 minimum trust to operate 1 2 3 4 with Darwin without continuous improvement weeks in production 1 new attack 2 model update 3 new rule 4 data drift ILLUSTRATIVE
New attacks, model updates, new rules, and data drift erode a static agent. Continuous improvement holds the line above the threshold your policy demands.
Read the trust-decay analysis →
WHERE IT FITS
spec YOUR INPUT discover DISCOVER verify DIAMOND deploy YOUR CI/CD defend DOME evolve DARWIN what production teaches amends the spec Dashed steps are yours, not ours
Darwin closes the loop of the Trusted Agent Lifecycle: evolve searches for a better agent before deployment, adapt hardens it in production. Every mutation is re-verified by Diamond before it ships.
  Vijil Discover   Vijil Diamond   Vijil Dome

Evolve your agent today

Free tier, no sales call. Point Darwin at your agent and review its first evidence-backed proposal.

Start for free