Research Scientist

Research Scientist

We're looking for a Research Scientist to set and pursue a research agenda for reliable long-horizon agents operating inside real enterprises.

Frontier labs optimize for general capability, and the public agent benchmarks are mostly sandboxes. Very little rigorous work exists on what it takes for an agent to reason across a system with nineteen years of undocumented decisions in it, plan a change across forty coupled steps, recover when step twelve reveals the model of the world was wrong, and be right often enough that a CFO signs the go-live.

Almost nobody has the landscapes, the traces, or the customers to study it. We do.

Two properties make this an unusually good research setting. First, much of the task space is verifiable — a transformation either produces a system that builds, passes regression, and behaves equivalently, or it doesn't. That's a real reward signal, not a preference model. Second, the parts that aren't verifiable are where the interesting work is: is this reconciliation correct, or merely plausible?

Was retiring that capability the right call? Designing reward and evaluation across that boundary is the central research question here.

You'll invent methods rather than only apply them, work with Research Engineers who help you run at scale, and hear from a product team within weeks whether you were right.

We'd like you to publish. Not everything, and never at the expense of shipping — but the work here is novel enough to be worth writing down.

One thing worth knowing up front: we post-train open-weight models on rented clusters and buy more compute when a result justifies it. We're constrained relative to a frontier lab. If your research only works at ten thousand GPUs, this is the wrong place.

• Set and pursue a research agenda on reliable long-horizon agentic behavior in real enterprise environments — you decide which questions matter, and defend the choice.
• Invent and validate methods for post-training agents on transformation work: reward design where verification is partial, delayed, or contested; RL formulations for long-horizon planning and tool use; curriculum and data strategy. Post-training and RL are the core of this role.
• Define how an agent remembers. Memory architecture for runs that span forty steps and days of wall-clock — what persists, how it's structured and retrieved, how it's revised when the world turns out to be different, and how a model is trained to use it rather than ignore it. This is one of the least solved problems in agentic AI and one of the most consequential for us.
• Own the question of what to measure. Develop evaluation methodology whose scores predict customer-observed correctness, and demonstrate where cheap automated proxies quietly fail.
• Study how multi-agent systems fail — error compounding across long trajectories, planning under partial observability of a landscape, delegation and verification between agents — and design against it.
• Work on the verification problem directly: how an agent, or another agent, establishes that a change preserved behavior when no test covers it.
• Solve how a system represents an enterprise to itself. Turning process, data, and code into an ontology or knowledge graph an agent can reason over reliably — and one that stays true as the underlying systems change — is a research problem we own rather than inherit.
• Investigate what post-training compute and data quantity buy us across model scales we can actually afford, and where the returns bend.
• Turn findings into things that ship, with Research Engineering and product.
• Publish papers, technical reports, and open-source artifacts, and represent Tessera's research externally.
• Raise the team's research bar: review experiment designs, mentor engineers moving into research, and be the person who asks whether the result is real.

• Designing a reward formulation for transformation work that doesn't collapse into reward hacking when the agent discovers it can pass the regression suite by removing the branch the tests don't reach.
• Developing an evaluation methodology whose scores track expert-reviewed correctness on changes no…