A control system for agentic work

Your AI agents ship fast.
You still have to prove it works.

Agents can claim. AVAA DEV proves.

AVAA DEV is neither an AI nor a model. It's a control system that puts AI agents to work and refuses to call their work "done" without verifiable proof — tests, gates, a sealed record. Verification is mechanical, not another LLM acting as judge.

Deterministic tests Independent verifier Cryptographic seal Local execution
avaa-dev — verifier payment.py
📁 src
  payment.py
  cart.py
📁 tests
  test_payment.py
📁 .avaa
  proof.seal
# generated by the agent — feature/payment
from gateway import bank
from avaa import seal
def validate_payment(amount, card):
if amount <= 0:
raise AmountError("invalid amount")
if not card.is_valid():
raise CardError("card declined")
receipt = bank.charge(card, amount)
return seal(receipt)
def refund(receipt_id):
receipt = bank.find(receipt_id)
return bank.credit(receipt)
▣ AVAA DEV — VerificationDONE AUTHORIZED
tests run — 142 / 142 passed
independent audit — promises = delivery
proof sealed — sha256 9f3a…c1d7
▣ AVAA DEV — VerificationBLOCKED · STATUS REFUSED
tests run — 138 / 142 (4 failed)
independent audit — promises ≠ delivery
proof — not produced · merge refused

AVAA DEV drives the AI agents of your choice (any provider), checks their work and seals the proof.

The problem

AI code ships faster than anyone can verify it.

The industry already has a name for it: verification debt. We generate at full speed — and trust doesn't keep up.

42%
of committed code is now AI-generated
96%
of developers don't fully trust AI-written code
48%
only verify it before shipping

Source: Sonar, State of Code — Developer Survey 2026 (1,100+ developers). "Verification debt" — term attributed to Werner Vogels, CTO of Amazon.

How it works

Five steps. Only one can write "done."

The agent doesn't decide it succeeded. The control chain decides for it — and only if every link holds.

1

Produce

The AI agent implements the task and announces "it's done."

2

Test

Deterministic tests and gates actually run.

3

Audit

An independent check re-reads: promises vs actual delivery.

4

Seal

A tamper-proof cryptographic record is produced.

5

Done

"Done" status authorized — only if everything held.

A link breaks? The status stays "not done." No proof, no "done." — that's our rule: NO CODE, NO DONE.

What it is

Agents produce. AVAA DEV verifies and proves.

An AI agent says "it's done." AVAA DEV doesn't take its word for it. It runs the tests, passes the gates, seals a tamper-proof record — and until the proof exists, the work is not "done." The producer never validates its own work.

Our thesis: governing agentic work means controlling it across three timesbefore (scope it, decide go/no-go), during (produce under control and prove it mechanically), after (harden: bugs, security, robustness). Existing tools cover only one. AVAA DEV targets all three — and at the moment that matters, the during, it proves instead of judging.

The demonstration

Same task. With and without control.

An agent announces "feature shipped ✅". On the left, you believe it. On the right, AVAA DEV demands proof before writing "done."

without-controlBLIND TRUST
$ agent run feature/payment
agent › implementation complete
agent › ✅ it's done
merge accepted (on its word)
deploying to production…
— 3 days later —
✗ regression in prod
✗ no proof of what was tested
? who validated it? the agent itself.
with-avaa-devPROOF REQUIRED
$ avaa run feature/payment
agent › implementation complete
agent › ✅ it's done → to be verified
gate › running tests… 142 passed
gate › independent audit… OK
gate › cryptographic seal… sealed
✓ verifiable proof produced
✓ status DONE authorized — not before.
who validated it? an independent check.

If a single step fails, the status stays "not done." (Illustrative demonstration — the real mechanism is reproducible.)

What changes

Four principles no one assembles.

Other tools record after the fact or have another LLM judge. AVAA DEV does the opposite.

Preventive

It blocks "done" without proof, instead of narrating the incident afterward.

vs. forensic log "we'll understand later"
⚖️

Separation of powers

The producer never validates its own work. Verification is independent.

vs. the agent declaring itself successful
🔏

Tamper-proof

Cryptographic seal + external verifier, against forgery of the proof by the producer itself.

vs. a ledger the producer can rig
⚙️

Mechanical

Deterministic gates and tests. Not another LLM that "judges" and can hallucinate in turn.

vs. LLM-as-judge
Proof, on ourselves

We don't just promise it. We hold ourselves to it.

AVAA DEV is built under its own control: every change to its code runs back through its own verification chain. These aren't customer numbers — it's the mechanism proving itself on itself.

8000+
automated tests on its own code — green suite, re-run at every gate
0
"done" status granted without a sealed proof — its own rule, applied to itself
3
separated powers — producer · decider · auditor; the producer never validates

We check ourselves the way we check everyone else: no critical security hole in our own code to date, and the few debts — dependencies to update — are tracked and fixed, never hidden. Cleanliness isn't a badge: it's the consequence of prevention.

Internal measures from AVAA DEV's own development (its test suite is re-run at every verification step). No customer data, no traction claim. Full traceability is shown in a reproducible demo to design partners.

▣ SEALED VERDICT — third-party verifiable
score without pipeline 5/5 fake-"done" · with pipeline 5/5 blocked
algo ed25519-code-state-demo-v1
code_hash 7c1659f3f8da…ff4fee
state_hash c193faef3035…46efeb
pub_key c218f02cb480…74c1d9
signature be7b8dec7e60f233…de4480f
Verify it yourself — no secret required, the public key is enough:
$ python avaa_seal.py verify
Signing needs the private key. Verifying needs only the public key above. One byte changed in the result, or a forged signature → TAMPERED. Reproducible demonstration, run locally.
The suite

Proof is the core. AVAA DEV is a full suite.

Everything essential to build software with governed AI agents — in one place, provider-agnostic and under control.

🏗️

Agents & Forge

Create, test, validate and activate your agents (agent Forge), with AI workshop, control plane and scheduling.

🔌

Models & providers

Provider-agnostic: routing, capability discovery, failover and resume, cost tracking.

🧠

Memory & context

Multi-level memory + semantic search, project twin and state persisted across sessions.

🛠️

Agentic tools

Computer use, scoped terminal and shell, browser, code tools, files, image generation, MCP, skills & connectors.

💬

Channels & surfaces

Desktop, CLI and a multi-channel gateway (Discord, Teams, WhatsApp, Signal, mail).

🔁

Software lifecycle

From bootstrap to deployment: project onboarding, integration & deployment, isolated execution environments.

⚖️

Governance & proof

Gates, sealed proof, delivery governor, evaluations, reviews, red team and audit — the core.

🛡️

Security & enterprise

RBAC, privacy (GDPR), security controls and observability.

Why now

The regulatory wind is blowing toward proof.

These texts do not mandate AVAA DEV and don't make it "required." But they make traceability of what AI produces increasingly expected in regulated sectors.

EU AI ActTransparency from Aug 2026; "high-risk" obligations pushed to Dec 2027 / Aug 2028. source ↗
DORAIn force since January 2025: operational resilience, ICT third-party register. source ↗
Cyber Resilience ActSBOM mandatory on 11 Dec 2027; vulnerability reporting from Sep 2026. source ↗
EU sovereigntyLocal execution: code and proofs that stay on your own infrastructure. source ↗

Timelines updated for the late-2025 "Digital Omnibus" agreement (still provisional until published in the Official Journal). AVAA DEV sells no compliance guarantee: it provides traceability and proof, not legal advice.

Who it's for

For whoever answers for what goes to production.

  • Security & compliance leaders (CISO) in regulated sectors
  • Finance & insurance subject to DORA
  • Public sector & operators sensitive to sovereignty
  • Teams deploying AI agents who must prove what they ship
  • Any developer who wants to be sure the generated code is actually implemented, not just announced
Data sovereignty

Your code and proofs never leave your infrastructure.

AVAA DEV runs locally, with no dependency on a third-party cloud: your data stays where you decide — on-prem, private cloud or sovereign region (EU, US, or other country).

↗ Sovereign cloud deployment (EU and other regions) coming soon.

Frequently asked

What people ask us most.

Is AVAA DEV an AI?
No. It's a control system that orchestrates AI agents and verifies their work. It does not generate the code itself: it proves the produced code does what it claims.
How is it different from a regular CI/CD pipeline?
CI/CD runs tests. AVAA DEV adds three things CI/CD doesn't have: separation of powers (the producer never validates its own work), refusal of the "done" status while proof is missing, and a sealed tamper-proof record verified independently.
Which AI models or agents are supported?
AVAA DEV is provider-agnostic and multi-provider: you plug in the models of your choice. It mediates between you and the configured LLMs, with no lock-in to a single vendor.
Does my data leave for a cloud?
Not necessarily. AVAA DEV can run locally: code and proofs stay on your own infrastructure, with no dependency on a non-EU cloud.
Is it available today?
AVAA DEV is in design-partner access. Access is deliberately limited: we co-build with a select few partners in regulated sectors, on real cases — not a promise, a reproducible demonstration.
Roadmap

What's coming next.

Around the already-live core — the during (mechanical proof) — the before and the after of control are being rolled out. Upcoming features — not shipped yet.

SOON
📦

Onboarding an existing repo — the before

Point AVAA DEV at an existing codebase: it reads it, generates the documentation suite (INDEX, ADRs, traceability matrices), then every change runs through the governed pipeline.

SOON
🧪

Hardening pass — the after

Between each phase and at the end of the run: hunting for unanticipated bugs, security analysis and robustness checks — so that behind the control, the code is close to flawless.

SOON
☁️

Sovereign cloud deployment

A sovereign cloud option in addition to local execution, for teams that want it — without giving up traceability or proof.

The pilot

In 2 weeks: your AI agents at work — and the proof of what they deliver.

We connect AVAA DEV to your context: your agents, the models of your choice (any provider), your tools, your code. You put them actually to work in one place — and everything they produce is checked and proven.

You walk away with:
  • your AI agents operational in AVAA DEV — driven in one place, memory and context kept from one session to the next, on the models you want;
  • real tasks taken end to end — work actually produced, not a toy demo;
  • for each one, the certainty of what holds and a proof verifiable by a third party of what the AI really delivered;
  • the same task shown both ways: shipped on trust vs shipped with proof — you see the difference with your own eyes;
  • an honest view of what remains uncertain — we only claim what we can prove.
Reserve a seat →
Design partner access

We're opening a few seats to build together.

AVAA DEV is in design-partner access. Access is deliberately limited: we co-build with a select few partners in regulated sectors, on real cases — not a promise, a reproducible demonstration.

or directly: contact@avaadev.fr

✓ Thanks, your message was sent. We'll get back to you shortly.