Agent = Model + Harness: why a production-ready harness is the real difference
When an AI agent fails in a real environment, the first reaction is almost always "we need a better model." That is the wrong diagnosis. In the vast majority of cases the problem is not the model — it reasons brilliantly. The problem is the harness: the software infrastructure around the model that turns it from something that answers into something that acts reliably.
This is the formula we build on:
Agent = Model + Harness
The same model, but a better harness — better results. This is where OpenKBS is strong: we build a production-ready, scalable, secure harness. This post explains what that means, why it matters for business, and how the two layers — the model and the harness — come together in the platform.
What a harness actually is
If the model is the brain, the harness is the body. The brain can think, but without a body it can't grasp anything, open a door, or check the result of its action. The harness is everything that sits between "the model decided what to do" and "the action happened safely, traceably, and repeatably":
- the tools the model can call;
- the isolated environment (sandbox) where code executes;
- the memory and storage that survive between sessions;
- the feedback loops that let the agent see the result and correct;
- guardrails — the boundaries that prevent the agent from doing harm;
- observability, which gives an audit trail for every action.
The model on its own is stateless — it takes text and returns text. Everything else that makes an agent useful in production is the harness.
Why most failures come from the harness, not the model
This is the key insight: most operational failures of AI agents come from the harness, not from the model itself. The typical symptoms have nothing to do with the model's intelligence:
- Context rot — the context overflows or gets polluted and the model loses the thread;
- Tool overload — too many, poorly described tools that the model gets confused between;
- Brittle wiring — hand-assembled integrations that break at the first change;
- Latency — every step goes through unnecessary network hops;
- Irrelevant retrieval — memory returns the wrong context;
- Weak verification — the agent doesn't check the result of its action;
- Missing guardrails — nothing stops the agent when it goes wrong.
None of these problems are solved by swapping the model. They are all solved with a better harness. That's why a strong harness makes average models useful, while a weak harness wastes even the best ones.
Layer 1 — The model: trust through zero data retention
We don't build models. And that's a deliberate choice: the best models change every few months, and being locked into a single vendor is a strategic risk. Instead we provide access to all the major providers — OpenAI, Anthropic, Google — through a single AI proxy hosted in our EU infrastructure.
The difference lies in the terms under which this happens:
- Zero data retention. Requests go through the providers under zero data retention agreements — nothing is logged, nothing is kept, and nothing is used to train models. Customer data does not stay with the provider.
- No API keys to manage. The customer doesn't juggle OpenAI, Anthropic, and Google keys — access runs through a single identifier and a single billing in credits.
- Supply chain consolidation. Instead of a separate contract, a separate risk assessment, and a separate audit for each AI vendor, the customer works with one provider. For regulated sectors this is a direct advantage under NIS2 and the AI Act — drastically fewer vendors to assess.
In other words: anyone can give you the models. We solve the part that really weighs in an enterprise environment — trust.
Layer 2 — The harness: this is where our strength is
A production harness consists of recognizable building blocks. The strength of OpenKBS is that each of them is implemented on top of managed, isolated, and certified infrastructure — not as a hand-assembled prototype, but as a platform.
| Harness building block | Implementation in OpenKBS |
|---|---|
| System prompts / context | Lambda functions — context and logic are code, versioned on every deployment |
| Tools | Project API: workers, S3, email, MQTT, database — ready-made tools the agent calls |
| Sandboxes (isolation) | Lambda microVM isolation + a dedicated AWS account per customer + on-demand EC2 workers for heavy tasks |
| Filesystem | S3 object storage with presigned URLs — limited by time and scope |
| Memory | Managed PostgreSQL (Aurora/Neon), point-in-time restore up to 35 days, 6 copies across 3 zones |
| Feedback loops | Agent loop in Lambda: tool_use → execution → observation → retry and correction |
| Guardrails | Multi-tenant isolation, injected secrets, credit limits, OWASP security audit, AES-256 / TLS 1.2+ |
| Observability | CloudWatch logs, worker logs, usage collector, administrative audit log |
| Model access | AI proxy — all vendors, zero retention, unified billing, no key management |
This is not a wishlist — it's the infrastructure that already sits beneath every project on the platform. The developer starts from a ready-made harness instead of assembling it from scratch for every agent.
Scalable by default
Lambda functions scale automatically from zero to thousands of concurrent executions. Heavy tasks (video processing, ML, batch) go to on-demand workers billed by the second. The database picks the right engine for the load. There is no capacity planning and no servers to maintain.
Secure by design
Every customer is physically isolated at the AWS account level — a hard boundary enforced by AWS IAM, not a logical separation. Secrets are injected at deployment, never in the code. Access runs through JWT and per-project keys with automatic rotation. Every new version goes through a structured security audit.
Production-ready means compliant
In a regulated European environment, "production-ready harness" carries one more meaning: a compliant harness. Here the strength of the harness and regulatory compliance merge into one.
- EU data residency — by default all resources are in the AWS eu-central-1 region (Frankfurt). Data does not leave the EU.
- Inherited certifications — the infrastructure builds on AWS with over 150 independently audited certifications (ISO/IEC 27001, SOC 2 Type II, BSI C5), including membership in Germany's critical infrastructure (KRITIS).
- Security audit on every version — static analysis for OWASP Top 10, checks for SQL injection, XSS, CSRF, SSRF, command injection, and CVE vulnerabilities before production; the reports are available to the regulator.
- Dedicated and transferable account — on contract termination the entire AWS account is transferred to the customer. Nothing is migrated, there are no proprietary formats, no vendor lock-in.
The same harness that makes agents reliable also makes them auditable. The details on NIS2 and the AI Act are covered in our separate posts — NIS2 and the AI transformation of the manufacturing sector and The AI Act and compliance for enterprises.
What this means for business
Most companies today aren't building a single AI agent. They're building dozens. And without shared infrastructure this quickly turns into agent sprawl — scattered, disconnected agents that no team can reliably manage, audit, or maintain.
A shared, production-ready harness solves exactly this:
- From demo to production. The prototype that runs on a laptop and the system that withstands real traffic in a regulated environment are two different things. The difference is the harness.
- Governance. One place for observability, secrets, limits, and audit — instead of every agent reinventing the wheel.
- Speed without taking on risk. Teams focus on the agent's logic, not on isolation, scaling, and compliance — those come ready-made.
The conclusion is simple: success in production requires harness engineering to be treated as a separate discipline, equal in importance to the choice of model. That is the discipline OpenKBS is strong in.
Next step
If your organization is moving from AI prototypes to real, production agents — especially in a regulated sector — get in touch. We'll look at your specific case and show you what a production-ready harness looks like for your business: isolated, scalable, secure, and compliant by default.
The enterprise services described — a dedicated AWS account, security audit, and review of AI-generated code — are part of the OpenKBS Enterprise plan.
This post is informational in nature and does not constitute legal advice.