Model degradation
The platform depends heavily on Anthropic’s Claude API. Model changes, deprecations, or performance regressions could materially impact product quality.
The Risk
AI models are not static. They’re versioned, fine-tuned, and occasionally deprecated. Our products wrap Claude’s capabilities:
- Murphy relies on Claude’s ability to analyse dependencies and predict delivery risks
- SmartBoxes uses Claude for code generation, file manipulation, and tool orchestration
- P4gent depends on Claude’s ability to understand supplier relationships and draft communications
If Claude’s capabilities regress—or Anthropic changes APIs in breaking ways—our products degrade immediately.
Specific Threats
- Model deprecation: Anthropic retires a model version we depend on
- Capability regression: New model versions are worse at specific tasks we rely on
- API changes: Breaking changes to tool use, context windows, or response formats
- Pricing changes: Cost increases that break our unit economics
- Rate limiting: Stricter limits that prevent our products from functioning at scale
Mitigations
Technical
- Model abstraction layer: Products don’t call Claude directly; they call our abstraction. Swapping models requires changing one config, not refactoring products.
- Prompt versioning: All prompts are version-controlled and tested. We can roll back to known-good configurations.
- Evaluation suite: Automated tests run against new model versions before we adopt them.
- Multi-model fallback: Critical paths can fall back to alternative providers (OpenAI, Google) if Claude becomes unavailable.
Business
- Relationship with Anthropic: Early access to model changes and deprecation notices
- Usage monitoring: Alert when model performance metrics drift from baselines
- Cost hedging: Unit economics assume 30% API cost buffer
Residual Risk
Even with mitigations, we cannot fully control Anthropic’s roadmap. If they pivot away from agent-native capabilities, we’d need to rebuild on different foundations. This is a concentration risk inherent to building on a single AI provider.
Probability: Medium (API changes are certain; breaking changes less so) Impact: High (could affect all products simultaneously) Mitigation effectiveness: Moderate (we can adapt, but not instantly)