What are this project's goals? How will you achieve them?

• Make interpretability research on/with agents faster and less painful — reducing the gap between having a hypothesis and testing it

• Enable researchers to bring their own techniques (SAEs, steering, activation extraction, etc.) into agentic workflows without rebuilding infrastructure each time

• Support replication and extension of existing interp papers by letting agents clone and operate inside paper repos directly

• Achieve this through a small, hackable open-source library built on Modal (for remote GPU sandboxes) and an IPython kernel execution layer for observability and iterative work

How will this funding be used?

• Dedicated development time to add features, improve stability, and expand provider support (already added Gemini and OpenAI alongside Claude)

• Modal compute costs for running and validating experiments

• Documentation and tutorials to help the interp research community onboard quickly

• Outreach to MATS scholars and other researchers to gather feedback and drive adoption

Who is on your team? What's your track record on similar projects?

• Primary author: Arya Jakkli, with 4 additional contributors (thienkhoi01, jnward, angkul07, mbiss10)

• Already shipped working case studies: hidden preference investigation, checkpoint diffing with SAEs, introspection experiment replication on Gemma 3 27B, Petri-style auditing agent

• 135 GitHub stars and 9 forks within ~5 months — organic traction from the interp community

• Built on and acknowledged by Goodfire (Scribe) and Modal; feedback from MATS scholars

What are the most likely causes and outcomes if this project fails?

• Low adoption: researchers stick with ad-hoc Claude Code + notebook setups; Seer's abstractions don't fit their workflows → project goes unmaintained, no harm beyond lost time

• Contributor burnout: small team, currently 1 primary author; without funding, development stalls before the library reaches a stable, widely-used state

• Infrastructure dependency risk: heavy reliance on Modal means if Modal's pricing or API changes, significant rework would be needed

• Superseded: a larger lab ships similar tooling with more resources → Seer becomes redundant, though early experiments and design decisions may still inform better tools

How much money have you raised in the last 12 months, and from where?

• No external funding raised to date

• Modal provided compute credits (new accounts receive $30 USD, likely additional credits given usage)

• This grant would be the first formal funding for the project