AI Agent Governance

Give every AI agent a sandbox, quotas, and a tamper-evident audit log

Runta wraps each agent in an isolated runtime with enforced resource quotas, a configurable egress allowlist, and an immutable log of every action taken. The governance package a security review expects.

runta-audit.log
$ runta exec --agent researcher --quota 512mb
[2026-07-14 09:12:01] SANDBOX_CREATED agent=researcher id=sbx_9f3a
[2026-07-14 09:12:04] EGRESS_ALLOWED host=api.openai.com rule=allowlist[0]
[2026-07-14 09:12:07] TOOL_CALL fn=web_search args={"q":"..."} tokens=142
[2026-07-14 09:12:09] QUOTA_WARN mem=438mb limit=512mb used=85%
[2026-07-14 09:12:11] EGRESS_DENIED host=s3.amazonaws.com rule=default_deny
[2026-07-14 09:12:22] AUDIT_SEALED hash=sha256:4c2e1a... immutable=true
[2026-07-14 09:12:22] SANDBOX_DESTROYED id=sbx_9f3a exit=0
$ _

Four governance primitives, one runtime

Every AI agent gets the same baseline: isolated execution, predictable resource use, controlled network access, and a complete tamper-evident record.

Sandbox Isolation

Each agent runs in a dedicated namespace. No shared state, no filesystem bleed, no access to sibling agent memory. Execution ends, the sandbox is destroyed.

Process-level isolation

Resource Quotas

Set hard limits on CPU time, memory, token throughput, and API call budgets per run. Agents cannot exceed their allocation. Overrun triggers a clean termination, never a silent hang.

Hard enforcement

Egress Allowlist

Network egress is denied by default. Operators define an explicit allowlist of hosts the agent may contact. Any attempt outside the list is logged and blocked, not silently dropped.

Default-deny network

Immutable Audit Trail

Every tool call, egress attempt, resource event, and lifecycle transition is appended to a sealed, content-addressed log. The log cannot be edited or deleted after the run completes.

Content-addressed

From agent call to verified record

Runta intercepts every agent action before it executes, enforces policy, and appends a signed entry to the audit log. Three steps.

01

Agent calls a tool

Your agent invokes a tool through the Runta SDK. The call is intercepted before execution. No changes to your agent code are required, only the runtime entrypoint.

02

Policy is checked and enforced

Runta evaluates the call against the active quota state, the egress allowlist, and any operator-defined policy rules. If the call is permitted, it executes. If blocked, the agent receives a structured rejection.

03

Immutable entry is sealed

Regardless of outcome, an entry is appended to the run's append-only log: timestamp, action type, inputs, outcome, and a content hash. When the run ends, the full log is sealed and signed. It cannot be altered.

Built to pass a security review

Runta surfaces the evidence a security team needs without requiring them to trust the agent's own reporting.

Network egress policy
ALLOW api.openai.com, api.anthropic.com, defined in allowlist
ALLOW your-internal-api.example.com, operator-approved host
DENY s3.amazonaws.com, not in allowlist, blocked and logged
DENY Any unlisted host, default-deny policy enforced
LOG All outcomes appended to sealed audit log

Tamper-evident logs

Each log entry carries a content hash chained to the previous entry. Any modification invalidates the chain. Verification is a single command.

Default-deny network

Egress is blocked unless explicitly permitted. Agents cannot exfiltrate data to arbitrary hosts. The allowlist is version-controlled and audited on every change.

Resource accountability

Every run records its peak memory, CPU time, token usage, and API call count. Budgets are enforced, not advisory. Overruns terminate cleanly and appear in the log.

Reproducible policy

Policy is declared in YAML and stored alongside your agent code. The policy version active at run time is captured in the log. What your security team reviews is exactly what ran in production.

Request early access

We are onboarding engineering teams deploying AI agents in regulated or security-sensitive environments. Tell us about your use case.

Ship governed agents from day one

Every agent your team deploys through Runta arrives with isolation, quota enforcement, and a complete audit record attached. No governance added later.