Respan, formerly Keywords AI Inc., today announced it has raised $5 million in seed stage funding to launch a proactive artificial intelligence observability platform that incorporates an automated evaluation agent to assist with prompt updates, send alerts and recommend fixes.

As AI agents become the norm across the AI landscape, observability and evaluation are increasingly necessary to keep them stable and on track. Familiar tools provide traces, metrics and AI-based evals to help teams inspect failures and measure quality.

“AI agents notoriously hallucinate,” co-founder Andy Li noted in an interview with SiliconANGLE.

Respan is entering a growing market for AI observability and evaluation tools, but Li argued that many existing platforms stop at retrospective tracing and scoring rather than helping teams proactively identify regressions, root causes and next steps.

To manage agents over time and with full visibility teams need tight knit observability, evaluation and decisions based on understanding agentic behavior that can handle non-determinism.

“The evaluation agent can analyze failure across different trials, find root causes to specific decisions agents are making, and recommend what eval to add next,” said Li.

Customers typically arrive at Respan with a framework, raw model, agentic setup or development environment already in tow. The platform integrates directly into whatever they already have with only a few lines of code, then surfaces traces of agent behavior in the platform within minutes.

Li said the company wants to position itself as model-, vendor-, framework- and language-agnostic. From Respan’s side of the observation, it doesn’t matter what model or code is running, the platform is still able to identify and trace agentic behavior and provide actionable, proactive alerts and recommendations.

The company uses off-the-shelf third-party models rather than proprietary foundation models for the AI under the hood. Customers can also choose their own model. Li noted, however, that when it comes to evaluation agents, it is best to pick a different model family from the one being used to run the agents being evaluated; otherwise, the bias of the watcher could interfere with the watched.

Similarly, he noted that customers should set their evaluation model to a “low temperature.” This means configuring it to produce more conservative, predictable and deterministic outputs. A low temperature forces the model to choose the safest choices, pushing it to behave more closely to its training data, so an instruction-following model becomes stricter.

Once wired up, the platform can suggest next evaluations, recommend changes, sample production traffic and fire alerts through Slack, email or text. It can also be aimed at development, staging or production so that teams can observe and evaluate agent behavior before turning agents loose in customer environments.

The round was backed by Gradient, Y Combinator, Hat-Trick Capital, Xiaoxiao Fund, Antigravity Capital and Alpen Capital along with prominent angels and AI founders.

Now that Respan has raised this funding, it intends to focus on hiring and infrastructure reliability. Over the past year, the company has scaled up to more than 100 startups and enterprise teams, bringing proactive AI observability insights to their agentic systems.

Image: SiliconANGLE/Microsoft Designer