Agents that improve themselves, and each other
Four new frameworks let agent teams evolve their organization, let foundation models improve without a teacher, and let vision-language models practice on their own generated data.
- In a preprint, EvoTeam jointly evolves team organization, interaction rules and individual agent skills, keeping a library of reusable teams built from execution experience. Self-reported m:…5327
- In a preprint, Recursa is a self-improving foundation model system that adapts to new domains by learning recursively from its own experience and validating improvements without human supervision, using a knowledge repository of skills, memories and reflections. Self-reported m:…3495
- In a preprint, SIMIT lets vision-language models improve on unlabeled queries by generating their own practice data, with gains across 17 benchmarks. Self-reported m:…2701
- In a preprint, MoLRSI is a recursive self-improving agent for molecular discovery that outperformed expert-led methods in antibiotic discovery and peptide-MHC binder design in testing. Self-reported m:…8161
Evolving the team, not just the agent
Most multi-agent research refines individual agents and hopes the team improves. EvoTeam, described in a preprint, instead treats the team as the thing to evolve. The collaboration-centric framework jointly evolves team organization, interaction rules and individual agent skills, using execution experience to maintain a library of reusable teams and context-specific interaction rules. m:…5327
The point is that collaboration across heterogeneous tasks is itself a skill. A team that worked for one task type may need a different organization for another, and EvoTeam aims to learn that mapping from experience rather than fixing the team structure in advance. m:…5327
Improving without a teacher
Recursa, described in another preprint, applies the same self-improvement idea to a foundation model system. It is designed to adapt to new domains by learning recursively from its own experience, validating improvements without human supervision. A knowledge repository stores domain-specific skills, task memories and reflections, which are refined through recursive self-improvement, while parallel learners execute tasks and validate improvements. m:…3495
DeepSeek-V4.1-Flash all-pass rate on Harvey LAB's Firm Knowledge benchmark rose from 30.0% to 55.0%. That matters because new domains rarely provide large amounts of expert-created training data, so a system that can learn from its own experience is not limited by what experts have already written down. m:…3495
Practicing on self-made data
SIMIT, described in a preprint, brings self-improvement to test time for vision-language models. The framework lets a model improve on unlabeled queries by generating its own practice data: it synthesizes diverse question, answer and image triplets, creating task-specific training examples without external annotations. Multi-stage verification and adaptive budgeting refine the generated data, and the framework achieved performance gains across 17 benchmarks. m:…2701
The verification step is the load-bearing part. Self-generated data is only useful if the bad examples are filtered out, and SIMIT's multi-stage verification plus adaptive budgeting, which allocates more samples to queries the model is less confident in answering, are designed to spend verification effort where it changes the outcome. SIMIT-FT achieved a +7.20% mean relative gain over BAGEL-7B. m:…2701
A self-improving specialist
MoLRSI shows what this looks like in a specialist domain. Described in a preprint, it is an AI agent for molecular discovery that automates and optimizes discovery workflows using persistent cross-task memory. It improves its search policy recursively based on performance feedback and refinement outcomes. m:…8161
In testing, it achieved a 15.9% experimental hit rate in antibiotic discovery against 8.7% for expert-led workflows, and produced eight experimentally validated hits against MART-1 and SARS-CoV-1 pMHCs. That result is the strongest claim in this cluster, and also the one that most needs independent replication: beating experts in a test setting is a starting point, not a deployment record. m:…8161
Across all four frameworks, the pattern is the same: the system keeps something from its own experience, a team library, a knowledge repository, generated practice data, or a search policy, and uses it to get better without waiting for a human to retrain it. The open question for builders is validation: a system that improves itself can also entrench its own errors, and the frameworks that win will be the ones whose checks keep up with their ambition. m:…5327 m:…3495 m:…2701 m:…8161
The reusable-team library in EvoTeam and the parallel learners in Recursa are the mechanisms that make self-improvement compound rather than reset. A team that is kept and reused, or a learner whose improvements are validated by other learners, carries its gains forward to the next task. Without that persistence, each task starts from scratch and self-improvement is just a label. m:…5327 m:…3495
What is still unresolved
- How do self-improving systems detect when they are entrenching errors rather than fixing them?
- Do the reported gains replicate outside the tested benchmarks and domains?
- What validation is strong enough to trust a system that retrains itself in production?
Provenance by Synorb. Every claim links to its source. This story was written by an agent from the Manifests listed below. Read the methodology and limitations, or send a correction.
Free to reuse, including commercially. Credit “Hanging Context, with provenance by Synorb” and link to this story. Keep the cited Manifest ids. Reuse terms v1.