We think the full breadth of human knowledge can, and should, be represented in natural language. Sphinx is a research lab measuring how far that holds true.

Scaling in-context learning

We’re scaling in-context learning (ICL) to allow AI to deeply understand how experts work. ICL remains under-explored at the scale of entire enterprises, domains, and professions. We’re building new algorithms to learn representations of knowledge, evaluating those representations by whether they enable agents to succeed on unseen tasks, and keeping them portable across agents and models. Our goal is AI that works with the judgment of an organization’s best people.

Better context means state-of-the-art performance at much lower cost

Our results show that ICL outperforms state-of-the-art fine-tuning techniques and enables generic ReAct loops to beat specialized harnesses, at substantially lower cost per task. These gains deliver materially higher ROI through better task success and lower inference costs. As we scale our training with more data and more compute, we expect these gains to compound across the full range of enterprise work.

ICL provides the interpretability and control needed for responsible AI

Our conviction in ICL goes beyond performance: its structural advantages make it ideally suited to the emerging demands of enterprise AI. ICL comes with enormous interpretability benefits.

Context is supplied explicitly at inference time, so we can trace individual agent rollout behavior all the way back to specific training data points. The causes of behavior are inspectable rather than opaque within model weights. We can inspect context after training, test how changes affect behavior, and deterministically incorporate human feedback.

In an environment of increasing concern around unsupervised agent behavior, ICL unlocks governance and control for context owners over how, where and when their agents get access to sensitive know-how. ICL also works with closed models and benefits directly from improvements in foundation models, carrying learned expertise forward as models improve.

Beyond prompt optimization to enterprise-scale learning

Prior work on methods like DSPy and GEPA demonstrated impressive results by optimizing prompts and examples for particular tasks. We’re scaling these ideas to significantly more ambitious targets: Sphinx uses ICL to learn entire enterprises, domains and professions. Fundamentally, this means moving away from optimizing fixed prompts to building a large-scale trainable context layer that can dynamically serve agents the right context at the right time.

The agentic ecosystem falls short at the context layer

Most agents can be summarized in three components:

• The model layer, expressed through a neural network.

• The harness layer, expressed through code.

• The context layer, expressed through natural language.

Each layer can be generic or specialized. At the model layer, builders can use frontier LLMs or opt to fine-tune an open-weight model. At the harness layer, general-purpose harnesses like Claude Cowork provide a strong baseline for autonomous work, while specialized tools like Harvey’s LAB harness for legal work and Devin Oncall for SREs push performance further within domains.

 Entire ecosystems support these specializations. We use Tinker or Spindle for fine-tuning and reinforcement learning. We develop harnesses with tools like LangChain or n8n, or buy specialized harnesses from vertical AI companies. 

The context layer is underserved

We think this is because context is too interpretable: since it comes in the form of readable natural language, people often treat it as documentation or configuration instead of a learned representation to optimize. Context building often amounts to taking notes, manually or with AI, or collecting metadata from systems of record. 

The result can look coherent and comprehensive, but readability is not evidence of effectiveness. The only question that matters is whether the context makes an agent better at its job. Naive context construction leaves substantial performance on the table; a more powerful learning algorithm turns the same enterprise data into measurable gains in agent accuracy, cost and ROI.

Sphinx makes context a trainable system

We emphasize the “learning” in in-context learning. We bring the same research ambition to ICL that has driven progress in parametric methods.

Sphinx develops algorithms to extract, learn, organize and  serve context. We train on large structured and unstructured datasets, generally sourced from the enterprises we work with so they reflect the reality of AI deployment in the wild. We study how performance scales with data and compute. We evaluate ICL methods by whether they enable agents to succeed on held-out tasks they couldn’t solve before, and by the resulting gains in cost and accuracy.

Solving AI’s last-mile problem

Our ambition is to solve AI’s last-mile problem: bridging general-purpose intelligence and the deep expertise needed to do real work autonomously. We believe ICL is structurally the right tool to do this scalably and interpretably.

Sphinx’s ICL lets AI approach unfamiliar problems with the understanding and judgment of a company’s most experienced employees. By synthesizing organizational knowledge, resolving contradictions, and learning continuously from outcomes, we aim to outperform any single human practitioner while preserving the transparency, steerability, and control required for consequential work.

Learning from real-world deployments

Sphinx is deploying advances incrementally, so from day one our platform for ICL has been essential to our work. Sphinx is live with customers and research partners, continuously validating and advancing ICL in real, commercially meaningful deployments. Our work is supported by investments from Lightspeed, Bessemer, SV Angel, and others.