Pure Python · 693 tests green

Governance for AI agents that decides, proves, and remembers.

Your agent asks before it acts. DarlaStack answers in microseconds, writes a tamper-evident record of every decision, and survives a crash with the audit trail intact.

agent.py
# the agent must ask the kernel before it acts
result = kernel.evaluate("send_email", {"recipient_count": 8})
# -> decision: allow   reason: allowed_rules_passed

result = kernel.evaluate("send_email", {"recipient_count": 5000})
# -> decision: block   reason: blocked_hard_rule
#    ...and an unforgeable line is written to the audit chain
~60 µs median per decision Tamper-evident hash chain Crash-durable signed audit Zero external dependencies to run
The problem

Autonomous agents act. You need proof they were governed.

An agent that can send, spend, or delete needs more than a prompt that says "be careful." It needs guardrails that actually block the dangerous call, a record of every decision that nobody can quietly rewrite, and that record has to outlive a restart. DarlaStack does those three things — and nothing it can't prove.

01 · DECIDE

It governs

Every action the agent proposes is checked against a frozen contract of rules. Safe calls pass; rule-breaking calls are blocked, with a precise reason code and a rule pass-ratio — in tens of microseconds.

02 · PROVE

You can't fake it

Each decision is appended to a SHA-256 hash chain. Change one past record and verification fails at the exact event. Spans are sealed under an Ed25519-signed Merkle root for court-ready evidence.

03 · REMEMBER

It survives a crash

The audit chain is persisted to durable storage. Kill the process and restart: the chain reloads, re-verifies, and the Merkle root returns bit-identical. The memory of what happened doesn't die with the process.

Live demo · real kernel

Try all three, right here.

This is the actual kernel running, not a recording. Send actions and watch them get allowed or blocked, forge a record and watch the chain catch it, then crash the agent and recover the proof from disk.

darlastack · live governance demo Open full screen ↗
Measured, not claimed

Real numbers from the test bench.

Every figure here was measured on the shipping code (Python 3.12, single thread). We show the honest production sizing, not just the flattering microbenchmark.

~60 µs
median decision latency (2-rule policy); p99 under 1 ms
500–1k
decisions/sec sized for realistic multi-rule production, per single core
~950 B
durable audit footprint per decision (JSONL on disk)
693
automated tests passing — kernel, persistence, and demo

Lab microbenchmark reaches ~12,000 decisions/sec single-threaded on warm, identical input; the 500–1,000/sec figure is the conservative number we'd size capacity against for real multi-rule workloads. Signed Merkle checkpoint over 1,000 events is ~2.5 ms. Crash recovery has been verified to reproduce the Merkle root bit-for-bit.

Honest scope

What it is — and what it isn't.

DarlaStack governs every action an agent submits, and now enforces: an approved action receives an unforgeable signed capability token, and a gateway runs the real effect only against a valid token — so a blocked action gets no token and cannot execute. Deploy the gateway behind a private boundary and it is the only path to the effect. It is not an OS sandbox; that you compose it with. Being straight about the boundary is how trust is earned.

Great fit

  • A fallible or misaligned agent that should be cryptographically prevented from costly actions
  • You need a tamper-evident, signable audit trail for compliance or forensics
  • The decision record must survive process restarts and crashes
  • You want guardrails as a library, with no external service to run

Needs more than the kernel

  • A truly adversarial process that bypasses the kernel entirely — pair it with OS-level sandboxing
  • Blocking actions the agent never submits for evaluation — the kernel governs what it's asked about
  • Distributed multi-node coordination — that's a layer above this kernel
Get started

Try it, or talk to us about licensing.

For developers

Run the kernel, the durable-audit layer, and this demo locally in two commands.

pip install -r darlastack_demo/requirements.txt
uvicorn darlastack_demo.server:app --port 8000
Launch the live demo →

For buyers

DarlaStack is available for licensing and evaluation. Tell us your agent's threat model and we'll map a deployment.

Request an evaluation