Public articles linked to the same research event.
arXiv The work introduces Stateless Language Agents (SLA): the harness owns the research state and reconstructs a fresh, role-specific context for every invocation, with a stateless Advisor assigning concrete experiments to parallel Workers from harness-summarized evidence; evaluated against EvoX, CORAL, and SwarmResearch on software engineering, kernel optimization, and algorithm design at budgets up to one billion cumulative tokens, SLA achieves the best final result on every task, matches the strongest kernel baseline's final performance with 93.1% fewer tokens (84.4% under Claude Code), and ablations show focused contexts and explicit assignments each contribute progress that can compound, while the Advisor consumes 0.24-0.51% of tokens.
The work introduces Stateless Language Agents (SLA): the harness owns the research state and reconstructs a fresh, role-specific context for every invocation, with a stateless Advisor assigning concrete experiments to parallel Workers from harness-summarized evidence; evaluated against EvoX, CORAL, and SwarmResearch on software engineering, kernel optimization, and algorithm design at budgets up to one billion cumulative tokens, SLA achieves the best final result on every task, matches the strongest kernel baseline's final performance with 93.1% fewer tokens (84.4% under Claude Code), and ablations show focused contexts and explicit assignments each contribute progress that can compound, while the Advisor consumes 0.24-0.51% of tokens.
The work introduces Stateless Language Agents (SLA): the harness owns the research state and reconstructs a fresh, role-specific context for every invocation, with a stateless Advisor assigning concrete experiments to parallel Workers from harness-summarized evidence; evaluated against EvoX, CORAL, and SwarmResearch on software engineering, kernel optimization, and algorithm design at budgets up to one billion cumulative tokens, SLA achieves the best final result on every task, matches the strongest kernel baseline's final performance with 93.1% fewer tokens (84.4% under Claude Code), and ablations show focused contexts and explicit assignments each contribute progress that can compound, while the Advisor consumes 0.24-0.51% of tokens.
The work introduces Stateless Language Agents (SLA): the harness owns the research state and reconstructs a fresh, role-specific context for every invocation, with a stateless Advisor assigning concrete experiments to parallel Workers from harness-summarized evidence; evaluated against EvoX, CORAL, and SwarmResearch on software engineering, kernel optimization, and algorithm design at budgets up to one billion cumulative tokens, SLA achieves the best final result on every task, matches the strongest kernel baseline's final performance with 93.1% fewer tokens (84.4% under Claude Code), and ablations show focused contexts and explicit assignments each contribute progress that can compound, while the Advisor consumes 0.24-0.51% of tokens.