§ 00 / THE ENFORCEMENT LAYER
System onlineEvery 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.
§ 00.2 — 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.
§ 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:
- Q. 01
"Walk me through who is allowed to call which tools, for which agents, and under which conditions."
- Q. 02
"Pull the audit log for every action this agent took on customer data in the last 90 days. I'll wait."
- 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
01 / intent
Agent decides to act
e.g. query database
02 / enforcement
Execlave enforcement
policy eval · audit · optional semantic check
p50 2.1 ms · p99 3.9 ms, excl. semantic check
03 / outcomes
- ✓Allowed — proceedsPASS
- ❚❚Paused — human reviewsHOLD
- ✗Blocked — stops hereDENY
04 / record
Audit log entry
hash-chained · signed checkpoints
Where enforcement fits
Enforcement is not logging, and it's not a prompt instruction.
| Capability | Execlave | Logging / observability alone | Prompt-level guardrails alone |
|---|---|---|---|
| Main job | Apply configured policy before a governed action runs | Explain activity, errors, cost, and behavior | Guide model behavior through instructions |
| Stop an unauthorized tool call | Yes — when the call is routed through enforce_policy() | Requires a separate enforcement layer | Instructions alone are not an external authorization boundary |
| Human approval | Policy can hold execution for review | Can capture or surface approval events | Can ask for confirmation, but needs an external workflow to enforce it |
| Audit evidence | Hash-chained audit trail sealed by signed checkpoints | Trace/log evidence; signing varies by implementation | Conversation history — not inherently a signed decision record |
| Best used with | Observability, evals, and tool permissions | Runtime enforcement | Runtime 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.
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.
§ 07 / Teams
One decision record. Three teams who need it.
Platform, security, and GRC teams share the same governed activity, with different questions to answer.
07.1
Platform engineering
Add the gate through Python or TypeScript, measure its latency, and keep a server-side stop available.
Govern agents with your platform team→
07.2
Security
Apply policy before tool calls reach systems; review injection, PII, and MCP integrity controls in your security workflow.
Review controls for security teams→
07.3
GRC and compliance
Follow decisions and human approvals into tamper-evident evidence that supports auditor questions.
Explore evidence for GRC teams→
§ 08 / Deployment
Run Execlave in our cloud. Or run it in yours.
Same product. Your choice of deployment.
08.1 / Managed
Cloud Managed for you.
- ▸Free tier available
- ▸Evidence exports for your SOC 2 Type II audit
- ▸Multi-region deployment on request
- ▸Managed Postgres + Redis
Start free→
08.2 / Self-hosted
Your infra Your data, your control.
- ▸Docker Compose or Kubernetes
- ▸Air-gap compatible
- ▸Agent and trace data stay in your network
- ▸Same product, your SLA
Explore self-hosted→
§ 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 RMFFramework 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.
01 / Permission
Which agent could request which action under the configured policy?
02 / Decision
Was the request allowed, held for review, or denied?
03 / Record
What tamper-evident record links the request, decision, and human approval when required?
Illustrative record — field names from the JSON audit export.
| action | policy.violated |
|---|---|
| resourceType · resourceId | policy · ‹policy id› |
| metadata.externalAgentId | support-bot |
| metadata.policyType | tool_invocation |
| metadata.enforcementMode | block |
| metadata.policyDecision.policyVersion | 3 |
| metadata.allowed | false |
| 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.
§ 12 / START
Talk through your deployment →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