Skip to content

§ 00 / THE ENFORCEMENT LAYER

System online

Every AI agent
your company runs
needs a gate.

§ 00.1 — Statement

Execlave is the enforcement infrastructure that sits between your AI agents and the real world — checking every governed action against your policy before it runs, and recording blocked and held actions in a tamper-evident audit trail.

Free tier. No credit card. Instrument an agent in minutes.

For platform engineering, security, and GRC teams. Choose cloud for a managed start, or self-hosted to run inside your network.

§ 00.3 — Ask your AIChatGPT ↗Claude ↗Perplexity ↗Grok ↗

§ 00.2 — At a glance

p50 server-side check
2.1 ms
Policy types
20
Frameworks
7
p99 kill-switch pause
5.4 ms

§ 01  / By the numbers

Server-side, local Postgres — methodology at /benchmarks

Execlave by the numbers.

01  ·  p50 server-side check2.1ms
02  ·  p99 server-side check3.9ms
03  ·  built-in policy types20
04  ·  compliance frameworks mapped7
05  ·  p99 server-side kill-switch pause5.4ms

§ 02  / The risk you can't see

Your AI agents are one prompt away from doing something you’ll regret.

An AI agent's output can reach your tools, APIs, and data directly. Without an enforcement layer between them, nothing checks an action before it runs.

Attack / Prompt injection

Your customer support agent receives a ticket:

The verdict / At the gate

Without an enforcement layer, nothing stops the agent from acting on that instruction — and an ordinary log won't show whether a policy allowed it. Your auditor will ask you to prove what happened.

Be ready for these three auditor questions:

  1. Q. 01

    "Walk me through who is allowed to call which tools, for which agents, and under which conditions."

  2. Q. 02

    "Pull the audit log for every action this agent took on customer data in the last 90 days. I'll wait."

  3. Q. 03

    "Show me the moment a human approved this agent's last 10 high-impact decisions."

Start with the permission, the decision, and the record.

§ 03  / Runtime enforcement

Runtime policy enforcement — a gate, not a log.

Every governed action is checked against your policies before it reaches the real world — in milliseconds, server-side.

// Execution flow with Execlave installed

request ────────▶ decision, before the tool runs

  1. 01 / intent

    Agent decides to act

    e.g. query database

  2. 02 / enforcement

    Execlave enforcement

    policy eval · audit · optional semantic check

    p50 2.1 ms · p99 3.9 ms, excl. semantic check

  3. 03 / outcomes

    • ✓Allowed — proceedsPASS
    • ❚❚Paused — human reviewsHOLD
    • ✗Blocked — stops hereDENY
  4. 04 / record

    Audit log entry

    hash-chained · signed checkpoints

Where enforcement fits

Enforcement is not logging, and it's not a prompt instruction.

Architectural categories — individual observability and guardrail products may also add enforcement features.
CapabilityExeclaveLogging / observability alonePrompt-level guardrails alone
Main jobApply configured policy before a governed action runsExplain activity, errors, cost, and behaviorGuide model behavior through instructions
Stop an unauthorized tool callYes — when the call is routed through enforce_policy()Requires a separate enforcement layerInstructions alone are not an external authorization boundary
Human approvalPolicy can hold execution for reviewCan capture or surface approval eventsCan ask for confirmation, but needs an external workflow to enforce it
Audit evidenceHash-chained audit trail sealed by signed checkpointsTrace/log evidence; signing varies by implementationConversation history — not inherently a signed decision record
Best used withObservability, evals, and tool permissionsRuntime enforcementRuntime enforcement and evals

Best results come from combining all three: enforcement at the gate, observability for context, guardrails for model behavior.

§ 04 / Interactive simulation — runs in your browser, fixture data

A browser-only simulation of the checks production agent calls go through: fire a prompt-injection attempt, watch it get blocked, then pull the kill switch and watch new actions stop. Nothing here calls our servers.

Try it · no signup
Live demo · runs in your browser
www.execlave.com/dashboard/halvorsen-freight · us-east-1simulated stream

Overview

Everything agents do, surfaced in real time.

Active agents

8

Traces · 24h

2,847

↑ live

Violations

12

Blocked

4

Live traces● streaming
  • ✓customer-support-v2118ms
  • ✓data-analyst342ms
  • ✓research-summarizer194ms
  • ✓code-reviewer221ms
  • ✓finance-reconciler88ms
No active alerts — enforcement running clean.

Simulated data. In production, violations and admin actions write hash-chained audit records

§ 05  / How it works

From zero to governed in minutes.

Three steps, minutes to ship.

01  /  Instrument

Add three lines to add the gate.

Install the SDK, then call enforce_policy() before the protected operation. Blocks, holds, or allows — your policy decides, before the tool runs.

02  /  Set rules

Declare what cannot happen.

Define policies: cost limits, model restrictions, content filters, prompt injection guards. Block or warn — your choice.

  • cost_limit · $50 / day
  • model · gpt-4o | claude-3.5
  • block · PII in prompts
  • guard · prompt injection

03  /  Ship with confidence

Deploy, watch, stop.

Deploy knowing every governed action is checked against policy, blocks and violations land in a hash-chained audit trail, and the kill switch is one click away.

Kill switch

server-side pause · p99 5.4 ms

§ 06  / AI Agent Management Platform

Beyond enforcement. The full control plane.

Execlave is an AI Agent Management Platform (AMP): the six controls enterprises need to manage autonomous agents in production — available today.

  • 01 / Capability

    Tiered autonomy governance

    Assign every agent an autonomy level — observe, advise, act-with-approval, or autonomous — and auto-apply the recommended policy bundle for that tier. Drift detection flags when an agent outgrows its guardrails.

  • 02 / Capability

    Real-time cost circuit breaker

    Synchronous spend caps per org, agent, user, or workspace across 1m/1h/1d/1mo windows — enforced in the policy path, not reconciled after the bill. Burn-rate alerts fire before a budget is breached.

  • 03 / Capability

    Agent registry & lifecycle

    Inventory every agent with a lifecycle state, immutable version history, diffing, and one-click rollback. Shadow-agent detection surfaces agents calling your API without registration.

  • 04 / Capability

    Permission-drift detection

    Baseline each agent’s tools, data sources, and permissions, then catch privilege escalation, anomalous access to sensitive or PII data, and unused over-privileged permissions.

  • 05 / Capability

    Eval-to-policy suggestions

    Turn detected anomalies, violation patterns, and eval failures into proposed policies with confidence scoring. Accept, modify-and-accept, or reject — nothing is ever auto-enforced.

  • 06 / Capability

    Data-access lineage

    Record what classes of data — public, internal, confidential, PII, PHI, PCI — each agent touched, with an auto-classifier, GDPR subject-access queries, and PII-by-agent reporting.

§ 09  / Compliance & security

Enterprise AI security and compliance — not an afterthought.

09.1

Row-level data isolation

PostgreSQL RLS ensures your data never touches another organization.

09.2

PII detection and scrubbing

Emails, SSNs, credit cards, and API keys detected and redacted before storage — on by default.

09.3

Immutable audit trails

Cryptographically verified, append-only logs. Tamper-evident by design.

09.4

Role-based access control

Principle of least privilege. Owner, Admin, Developer, and Viewer roles across the platform.

Evidence mapped to

SOC 2 Type IIHIPAAGDPRISO 27001EU AI ActPCI DSSNIST AI RMF

Framework mappings help organize governance evidence. They do not certify your organization or guarantee compliance.

§ 10  / Evidence

Answer an auditor with the decision and its evidence.

Execlave records blocks, policy violations, and admin actions in a hash-chained, append-only audit trail sealed by cryptographically signed checkpoints. Use that evidence to investigate activity and support control reviews.

  1. 01 / Permission

    Which agent could request which action under the configured policy?

  2. 02 / Decision

    Was the request allowed, held for review, or denied?

  3. 03 / Record

    What tamper-evident record links the request, decision, and human approval when required?

Illustrative record — field names from the JSON audit export.

actionpolicy.violated
resourceType · resourceIdpolicy · ‹policy id›
metadata.externalAgentIdsupport-bot
metadata.policyTypetool_invocation
metadata.enforcementModeblock
metadata.policyDecision.policyVersion3
metadata.allowedfalse
contentHash‹SHA-256, computed at write›
previousHash‹contentHash of the prior record›
chainValid‹result from /chain/verify›

Illustrative values, not a live export. Real exports (GET /api/v1/audit-logs/export) carry the computed hashes; check a chain with GET /api/v1/audit-logs/chain/verify.

§ 11  / FAQ

Questions about AI agent governance.

Understand the enforcement boundary, deployment choices, and evidence before you integrate.

Q. 01How fast is the enforcement layer?

Server-side policy checks measure p50 2.1 ms and p99 3.9 ms, including the database round-trip, against local Postgres in a single sequential worker. That excludes HTTP authentication, network time, and model-backed semantic checks — see /benchmarks for the method. Over the network your client adds round-trip time, and your agent's user-facing latency is usually dominated by model inference.

Q. 02Can I run Execlave on my own infrastructure?

Yes. Self-hosted deployment is available from the Starter plan upward; the Free plan is non-commercial cloud evaluation only. You get the same Docker images we run in our cloud, plus a JWT license key. Agent and trace data stay on your infrastructure; the license check sends usage metrics only, and air-gapped mode sends nothing. See the self-hosted page for details.

Q. 03Which compliance frameworks does Execlave cover?

SOC 2 Type II, ISO 27001, EU AI Act, GDPR, HIPAA, PCI DSS, and NIST AI RMF. Compliance reports are generated from your actual audit log — not from a questionnaire — and map to the specific controls each framework requires (CC6.1, A.9.1, Art. 14, Art. 30, §164.308, etc.).

Q. 04How is this different from Guardrails / Lakera / Promptfoo?

Those tools focus on input/output validation at the prompt layer. Execlave is the governance layer one level deeper: it sits between the agent and the systems it can act on (tools, APIs, databases, files). Prompt validation catches some attacks; tool-level enforcement blocks unauthorized actions routed through the gate regardless of which prompt produced them, and produces the audit trail your compliance team needs. They are complementary, not competing.

Q. 05Is Execlave an AI Agent Management Platform (AMP)?

Yes. Beyond runtime enforcement, Execlave provides the full agent-management control plane: tiered autonomy governance, a real-time cost circuit breaker, an agent registry with lifecycle and versioning, permission-drift detection, eval-to-policy suggestions, and data-access lineage. All are on by default in both the managed cloud and self-hosted deployments.

See all FAQs

Your agents are already running. Start governing them. — Execlave: AI Agent Governance and Enforcement Platform

Free tier. No credit card. Instrument an agent in minutes.

No credit card required