Policy on every agent call and nine services in production: an enterprise AI control plane for a UK software vendor

Client context

swarmd.ai sells a control plane for AI agents to enterprises in regulated industries. It sits between agents, tools and model providers, and enforces policy, audit, identity and spend controls on every call.

What we built

We owned the engineering end to end, from architecture to production.

• Policy on every action. Customer policies are versioned and applied from tenant level down to a single agent, so the rule in force at any moment can be shown to an auditor.
• A tamper-evident audit trail. Every call is recorded with its source, destination, outcome and timing, in a log that shows if it has been altered.
• Enterprise identity. Per-tenant single sign-on against each customer’s own identity provider, central credential rotation and encryption of sensitive data.
• Tenant isolation. Enforced in the database as well as in application code.
• Guardrails. Sensitive-data detection, rate limits, spend controls and human approval for consequential steps.
• Multi-protocol, multi-provider. Several agent protocols, the major model providers and a payments rail behind one control plane. Adding another is an addition, not a rewrite.

Built to scale

• Container-native services on Kubernetes that scale out independently as load grows.
• Rolling, feature-flagged releases and zero-downtime database migrations.
• Logs, metrics and traces correlated across every hop, in place before each feature went live.
• Automated quality gates, security scanning and end-to-end tests on every release.

The codebase is now run by swarmd.ai’s own engineers.

Results

See the platform at swarmd.ai.