Morph launched Reflexes, an API that delivers semantic signals from AI agent traces in under 30 milliseconds, targeting companies running production agents at scale.

The system addresses behavioral failures in AI agents — looping, reasoning leakage, and user frustration — that frontier models like GPT-4 or Claude are too expensive to monitor continuously across millions of agent interactions.

Multi-head architecture cuts inference costs

Reflexes uses a custom inference engine forked from vLLM that reuses 99% of prefill computation across multiple classification tasks. The architecture removes the decode step from a modern LLM and employs hybrid attention with shared KV cache across all classification heads.

Running four reflexes versus 100 adds less than 2ms overhead, according to the company. The full API request completes in under 90ms regardless of the number of behavioral signals being tracked.

"If you're even a medium-sized startup, you're dealing with tens of thousands of agent runs and millions of turns," said Tejas Bhakta, who built the system. "Frontier LLM-as-judge does not scale."

Bhakta previously developed similar infrastructure at Tesla for sampling signals across petabytes of autonomous vehicle data.

Production focus over dashboards

Morph positions Reflexes as API-first tooling rather than another analytics dashboard. Companies can train custom reflexes through the web interface, then deploy them to trigger automated responses to specific agent behaviors in production.

The system targets engineering teams monitoring behavioral patterns like user frustration rates across large-scale agent deployments where per-turn evaluation with frontier models becomes prohibitively expensive.

Reflexes builds on multi-head classification techniques from 2019-era BERT architectures, adapted for modern transformer attention mechanisms and production inference requirements.

Morph offers the service through its existing platform, which includes agent development tools. The company has not disclosed pricing for the Reflexes API.