Epi Represents a Step Toward a Persistent, Self-Improving Agent for Autonomous Research

Meta description: Épi, an autonomous research agent, sustains multi-session investigations and produced 120 record-setting results across mathematical bounds and coding tasks.

Tags: Épi, autonomous research agent, mathematical records, Épi-turbo, FrontierOR

Featured image: An abstract technology illustration of persistent memory nodes, branching timelines and independent verification paths, with no real people, no company logos and no text.

Épi Autonomous Research Agent Produces 120 Record-Setting Results

Épi, an autonomous research agent, has produced 120 record-setting results by carrying investigations across model sessions and evolving its methods through accumulated experience. The reported output spans two modes of progress: tightening numerical intervals and producing construction-level results. It includes 20 numerical bound records across 12 families and 100 construction records across eight families, and it also extends to executable algorithmic solvers through Épi-turbo.

Background: A Need for Research Continuity

Many model-assisted workflows stop when a session ends. Hypotheses, failed attempts and partially explored branches can disappear, forcing researchers to restate the problem. Épi is presented as a response to that discontinuity. The agent combines persistent research history, hierarchical memory, branching and independent review so that it can keep investigations continuous, improve through experimental feedback, and reopen alternative paths.

Development: Sessions, Memory and Verification

The timeline begins with continuity: the agent lets experiments run across session boundaries rather than confining them to a single conversation or model call. A fixed verifier evaluates their outputs, giving a stable checkpoint for what counts as a successful result. Because the verifier is fixed, changes in memory or branching do not redefine success midstream. That makes session-spanning work auditable: earlier outputs can be checked against the same standard.

Results: Numerical Bounds and Construction Records

The reported results combine quantitative tightening and structural construction. Épi produced 20 numerical bound records across 12 families and 100 construction records across eight families. The numerical improvements include an approximately 84-fold narrower interval for the n-queens constant and a more than 3.6-million-fold narrower rigorous interval for the quartic Feigenbaum attractor dimension. Exact dual certificates also establish that the Cohn-Elkies linear-programming bound cannot be sharp in dimensions 10 and 11.

Beyond numerical improvements, Épi gives a computer-assisted proof of Tuza's conjecture for split graphs with an eight-vertex clique part and at most three triangle-active neighborhood types, with arbitrary multiplicities. The record counts are presented by family, framing the output as breadth across problem groups rather than one isolated improvement.

Extension: Algorithmic Coding with Épi-turbo

The same research process moves from proof-oriented records to executable solvers through Épi-turbo, a lightweight configuration for building solvers for tasks including vehicle routing, scheduling and production planning. On 30 FrontierOR tasks, Épi-turbo achieves a 5.4% higher total benchmark score than Codex at 64% of its model cost, using the same underlying model and reasoning effort.

The current stage is a demonstrated workflow rather than a single result list: persistent sessions, verification, and interpretable outputs are connected through a system designed to aid human understanding by tying computational results to explicit methods and mathematical arguments that researchers can inspect and extend. What to watch next is how such session-spanning verification and memory support longer, more complex research workflows.