AI Agents

This paper frames memory management not as a fixed architecture, but as a trainable skill. AutoMem introduces a dual-loop system. The first loop has a powerful LLM meta-review agent trajectories to iteratively improve the memory structure (file schemas, action vocabulary). The second loop identifies and reinforces the agent's good memory decisions, fine-tuning its proficiency. On long-horizon tasks like NetHack, optimizing memory alone boosted a 32B model's performance by 2x-4x, making it competitive with much larger frontier models.

This paper introduces RepoRescue, a benchmark for the practical task of updating old repositories to run in modern environments. The setup is realistic: an agent is given a repo that passed its tests historically but now fails, and it must produce a patch. The evaluation is rigorous, including a runtime-enforced mode that blocks agents from editing the test files. Under these conditions, Kimi still rescues 41.5% of Python repos. The study also finds different agent systems are complementary, suggesting an ensemble could be highly effective.

Standard software engineering agents can be misled by bug reproduction tests (BRTs). This paper introduces SWE-Doctor, which finds that naively using BRTs leads to partial or incorrect fixes. Instead, SWE-Doctor generates multiple BRTs for different aspects of a reported issue, executes them to create runtime diagnosis records, and then feeds these diagnoses to the patch generation agent. This runtime grounding provides a much richer signal, significantly improving bug-fixing performance on SWE-bench across five different agent backends.

How do you trust a workflow when any step can be an LLM hallucination? This paper proposes Agentic Transaction Processing (ATP), a model implemented in the Mnemosyne runtime. It treats all agent-generated actions as untrusted proposals that are only committed to state after passing a set of deterministic, executable constraints. This separates the competence of the proposing agent from the correctness of the committed state. The system includes proven safety guarantees and a protocol for localized, bounded reactive repairs.

This sharp audit questions the reliability of popular performance-optimization benchmarks like GSO and SWE-Perf. The authors found that reference patches often fail their own validity rules when replayed on different cloud machines, with SWE-Perf being especially fragile. Leaderboard rankings are also unstable, sensitive to arbitrary scoring rules that can give huge weight to a few tasks. Perhaps most damning, at least one public submission already matches or beats the official reference patch on 85.3% of the tasks analyzed.

In an interview, Introspection co-founder Roland Gavrilescu discusses "autoresearch," the company's approach to building self-improving agent systems. The core idea involves creating agent "recipes" that define tasks, which are then refined through feedback loops. While the goal is increased automation, Gavrilescu emphasizes that humans remain central to defining the high-level goals and architecture of these software-producing agent factories. The piece offers high-level patterns rather than deep technical specifics.

This interview with Cursor's Pauline Brunet outlines their "Forward Deployed Engineer" model for enterprise agent adoption. The FDE team acts as a bridge, helping companies set up internal "software factories" powered by AI agents. They identify high-value use cases, integrate with existing dev workflows, and essentially product-manage the initial agent deployments. The approach mirrors Palantir's model, focusing on high-touch services to solve the last-mile problem of getting agents to work inside complex organizations.