Boidra Platform

Agents make decisions you can’t see. When something goes wrong - or right - you have no trace, no score, and no way to compare runs. Standard APM stops at the HTTP boundary; it doesn’t understand a reasoning step.

The SAP-native options don’t close the gap yet. SAP Cloud Logging exists, but it wasn’t built to trace and score reasoning steps, and AI-Hub-level observability isn’t fully available. That leaves teams either flying blind, or shipping sensitive telemetry off to third-party tools.

Meanwhile the questions get harder. Finance wants to know what the agents cost. Audit wants to know who changed what. And nobody can say whether last month’s prompt change made the output better or worse.

One agent is easy to reason about. A flock is not - and production is always a flock: agents calling tools, calling models, calling each other across systems. You cannot operate what you cannot see, and you cannot improve what you never scored.

So the platform starts from a single unit: the span. Every step an agent takes - a model call, a tool invocation, a handoff to another agent - emits one. From that one primitive, everything else follows. Stack spans into a trace and you can debug. Price them and you have FinOps. Sign them and you have an audit trail. Score them and you have evaluation.

That is the whole design: instrument once, then answer the operational, financial, governance and quality questions from the same data - instead of buying four tools that each see a fraction of the picture.

Four surfaces over one stream of spans:

• Observability - a span for every agent step, with the full trace waterfall: tools called, inputs, outputs, latency and errors. SLOs and budgets with live burn-down, and alerts when error rates or token spend cross your thresholds.
• FinOps - cost in euro and Capacity Units, broken down per agent, per project and per model. Usage details and scheduled reports, so finance gets an answer without asking engineering.
• Governance - an audit trail of who ran what, when, and what it changed. Transports to promote agent configuration across dev, test and production, so a change is a reviewable event rather than an edit in a console.
• Evaluation - LLM-as-judge scoring against criteria you define, with eval datasets built from real production traces. Catch a regression before your users do.

It runs as a BTP-deployed app, so observability lives where your business already runs - no context-switching to an external dashboard.

Instrumentation is a dependency and a destination, not a rewrite. You add the boid agent to your service, point it at your Boidra Platform instance, and spans start arriving on the next request - no change to your business logic and no proprietary SDK to code against.

• Install - add the agent package to your Python or Node service. It wraps the LLM and tool clients you already use.
• Configure - one destination and one API key, supplied through the BTP destination service rather than hardcoded. Nothing else to wire.
• Deploy - push as normal to Cloud Foundry or Kyma. The agent ships with your app; there is no separate collector to operate.
• Verify - the first trace appears in the console within seconds, and the service shows up in the estate view automatically.

Because it is built on OpenTelemetry, anything already emitting OTel spans can report in without our agent at all - point your existing exporter at the platform endpoint. A2A and MCP calls are traced end to end, so a handoff between two agents stays a single connected trace rather than two unrelated ones.

Agents we build for you arrive instrumented from the first prototype. For agents you already run, expect an afternoon rather than a project.

The platform is native to where your agents actually run: deployed on SAP BTP, wired to Cloud Foundry and your existing destinations, and speaking the open protocols agents use - OpenTelemetry for spans, A2A and MCP for the calls between agents and tools.

That means instrumentation is not a rewrite. Drop the agent in, and spans start arriving - whether the agent is a custom Joule skill, a Gen AI Hub application, or an A2A flock you built yourself.

This is the part that matters most. Boidra Platform keeps all telemetry inside your own SAP BTP landscape - your data never leaves your organization. We do not export traces, prompts or outputs to third-party observability platforms like MLflow or any external SaaS.

For teams handling sensitive business data, that data sovereignty is not a nice-to-have; it is the requirement. It is also why the platform is deployed into your account rather than sold as someone else’s cloud.

Small and medium teams running custom AI inside SAP - the ones without a dedicated AI operations department, who still have to answer for what their agents do, what they cost, and whether they are getting better.

We build it because we need it. Every agent Boidra ships for a customer is developed, evaluated and operated on this platform, from the first prototype through to production.