When swarmd.ai set out to build a new multi-agent orchestration platform, the software company engaged Hexploits to help engineer the platform while effectively managing the risks at play.
“swarmd.ai represents what we believe the next decade of enterprise software looks like – autonomous agents doing real work, mediated by control planes that make them auditable, safe, and accountable,” notes a statement from Hexploits’ website. “We built theirs end-to-end in six months, at 75% under budget, on the patterns and practices that have served us across every engagement. Enterprise-grade orchestration for the agentic web.”
The challenge, when Hexploits arrived on the project, was that the agentic ecosystem is moving faster than the standards underneath it. New protocols ship monthly, and specs that were experimental last quarter are production assumptions today. Customers want to connect agents from three vendors, route them through two model providers, and audit the whole graph for compliance, immediately.
With swarmd.ai looking to build the control plane “the agentic web has been missing”, the firm believed its answer was a tenant-aware control plane that sits between agents, tools, and language models, enforcing policy on every hop and capturing an immutable record of what each agent said and did. Having contacted Hexploits with “a vision, a deck, and a deadline”, the consultants set to work.
Treating early-stage platform builds as exercises in disciplined scope, ruthless reuse, and engineering practices that compound, Hexploits recognised the faster its team could ship a change safely, the more change they could ship inside a fixed budget. This saw the professionals quickly lay the platform out as a modular monorepo, where every protocol contract, every cross-cutting concern, and every bounded context lives in its own module.
Policy as code
According to Hexploits, “Customer policies are written in a purpose-built policy language and evaluated by a battle-tested policy engine - the same approach Fortune 500 security teams already trust. Bindings reference an immutable policy version, so an audit replayed in eighteen months evaluates against the exact rules that ran on the day.”
The result is policy that compliance teams review the way they review infrastructure – “not a black box, not a backend developer’s secret.” This yields a level of audit that a regulator will actually sign off on – featuring transactional outbox delivery so business writes and audit writes succeed together or not at all; hash-chained, tamper-evident events so any tampering breaks verification at the next checkpoint, with regulator-grade timestamping on the roadmap; a hot operational store for live queries and a warehouse for analytics so "every call from agent X in the last quarter" runs in seconds, not minutes; and correlation IDs that survive every hop so a single trace can be followed end-to-end without re-joining logs by timestamp.
At the same time, a single mistake in multi-tenant SaaS is one customer’s autonomous process getting another customer’s credentials. The bar is higher than traditional B2B, not lower.
To meet these standards, Hexploits integrated with a mature identity broker, wired up per-tenant SSO for enterprise customers’ own identity providers, and built credential rotation as a first-class operation: rotate once, enforced everywhere. Sensitive material in the system is encrypted at the application layer before it ever touches a database, so even a stolen backup yields nothing useful.
This means zero trust is not a checkbox. It is a posture – applied consistently at every layer, or not at all. At the network, every service-to-service hop is authenticated. No service trusts another by virtue of sharing a subnet, a cluster, or a load balancer. Tokens are scoped, audienced, and short-lived; a compromised service does not become a master key.
At the application. Every request is authenticated and authorized at the boundary it enters. Tenant scope is enforced in code, not by convention. The principle of least privilege is the default, not the goal. At the database. Sensitive material is encrypted at the application layer before it ever lands on disk. Each service connects with its own credentials and reads only what its role permits. Row-level isolation by tenant is enforced inside the database, not just outside it.
The result is an attack surface that does not collapse when one assumption is wrong. A stolen backup yields nothing useful. A compromised service cannot pivot. A bug that bypasses one layer of authorization still hits the next.
Middleware that delivers
Every request that crosses the platform flows through a typed, ordered pipeline: skill restriction, sensitive-data detection, rate limits, human-in-the-loop holds. The pipeline composes cleanly, short-circuits cleanly, and is type-safe end to end - the compiler refuses to build a change that forgets a case. New behaviour is a new pipeline step, slotted in by configuration.
Meanwhile, tech choices “stayed boring on purpose”. The unifying principle: “pick the technology your customers’ platform teams already run, then make it sing.”
This helped to produce a complete enterprise platform, delivered in six months, at 75% under budget and 75% ahead of the original timeline. More importantly, it was the platform swarmd.ai needed to take into enterprise sales conversations - with the engineering integrity those customers’ compliance teams demand.
The handover bar was clear: “a swarmd.ai engineer joining on Monday should be productive by Friday” – something Hexploits says it hit without issue, while ensuring patterns are consistent across every service, and tests run against real dependencies. “The platform reads like one built by a team that intended to live with it – because that is the standard we hold ourselves to… Affected-based pipelines, automated quality gates, end-to-end tests on every release, and pre-commit checks engineers feel before code even leaves their laptop. The team ships small changes, often, with confidence.”
Results
swarmd.ai walked into enterprise sales conversations with a platform whose answers to the hard questions - audit, isolation, policy, rotation, compliance posture - are concrete, demonstrable, and built.
For customers of swarmd.ai, the promise of agentic AI in the enterprise has always been gated on governance. Customers can now plug any compliant agent, any compliant tool, and any major model provider into swarmd.ai and get a single audit trail their auditors will accept; policies their compliance team can read; credentials they rotate once, enforced everywhere; rate limits and human review they configure, not negotiate; and an identity model that respects their existing identity provider, not one that demands replacement.
Looking back on the project, Hexploits concludes, “We measure success by our clients’ success. For swarmd.ai, that means: a platform that already does what the sales deck promised, in production; an architecture that absorbs new protocols and new compliance regimes without rewrite; a codebase their growing team can extend confidently, without us; and 75% of the original build budget still on the balance sheet, funding the next year of growth.”
“This is what we mean by AI engineering done seriously. Not a wrapper around a model. A platform that lets the models be useful, safely, in companies that have rules.”