Agents are starting to keep what they learn, and share it with the next agent
Zep turns verified agent runs into reusable skills, Nous turns employee workflows into shared organizational skills, Legora splits skills from subagents, and Meta's agents now propagate techniques across its model fleet.
- Zep Agent Skills compiles verified agent execution traces into reusable procedural memory, and DeepSeek V4.1 Flash with skills at low effort achieved a 56.8% pass rate on EnterpriseOps-Gym, above the 52.8% score of Claude Opus 5.5, according to Zep. Self-reported m:…5622
- On a separate benchmark, tokens per passed run fell by 7.0% to 44.6% in all eight configurations when using skills, and Zep scans skills for credentials and injected instructions, according to Zep. Self-reported m:…5622
- Nous Research's Hermes Agent is now available in Microsoft Teams, its Collective Wisdom turns successful employee approaches into reusable shared skills, and the enterprise version runs inside customer-managed Azure tenants, according to Nous. Self-reported m:…4352
- Meta says its Ranking Engineer Agent handles over 60% of technique propagation across its ads ranking models as of March 2026, engineering productivity has increased 3 times, and it estimates over $45 million in compute savings by the end of 2026. Self-reported m:…4731
Skills learned from doing the work
Agents in production often lose the knowledge they gain from solving problems, because successful runs do not change model weights or prompts. Zep has introduced Zep Agent Skills, a system that converts verified agent execution traces into reusable procedural memory by compiling successful task runs into versioned files in the open SKILL.md format. Procedures are learned only from verified successes confirmed by a registered tool, an authorized person, or an external system, which is the design choice that matters most: the skill comes from work that was checked, not from any run that happened to finish. The idea is that a cheaper model can then use the performance of a stronger teacher model without anyone retraining model weights. m:…5622
The numbers Zep reports are striking, and they are the company's own results on its chosen benchmarks, so they should be read as a claim to test rather than settled fact. DeepSeek V4.1 Flash with skills at low effort achieved a 56.8% pass rate on EnterpriseOps-Gym, exceeding the 52.8% score of Claude Opus 5.5. Skills raised the number of tasks that pass in all four runs in seven of eight configurations on a separate benchmark, and tokens per passed run fell by 7.0% to 44.6% in all eight configurations when using skills. Zep applies security controls including scanning for credentials and injected instructions, and namespace isolation for skills. m:…5622
From one employee's fix to the whole organization
Nous Research is applying the same idea at the level of the organization. The company has partnered with Microsoft to bring its Hermes Agent to the enterprise, with Hermes Agent now available in Microsoft Teams and Hermes Desktop available in the Microsoft Store. Its Collective Wisdom feature allows employees to turn successful problem-solving approaches into reusable skills: a successful approach becomes a written definition paired with code, and that shared intelligence remains a private asset of the organization. The enterprise version allows organizations to run agents in their own Azure tenant, and Hermes Agent workloads run in Azure Container Apps Sandboxes, which addresses the security and data residency requirements that usually block this kind of sharing. m:…4352
Nous also says Hermes Agent ranks number one among public open agents on OpenRouter and processes nearly 2 trillion tokens a day. An agent that processes that much work generates an enormous number of traces, and the value of Collective Wisdom grows with the volume of successful work it can learn from. The risk runs the same way: a bad approach shared widely is worse than one tried once, which is why verification matters. m:…4352
Skills are not subagents, and the fleet learns too
Legora draws a distinction that clarifies the whole trend. The company uses Skills to capture and deploy legal team standards as reusable instructions, and subagents for tasks that benefit from separate execution, such as parallel legal research. Adding expertise and dividing work are distinct design decisions, and Legora evaluates the benefits of separate execution against coordination costs, time, resources, and review requirements. A team can maintain and reuse Skills like agreement-review processes or client-specific guidance across different assignments, and applying a new area of expertise does not inherently require a separate worker. Subagents receive a task and relevant context to return focused findings for the main agent, but delegation requires a brief and a mechanism to incorporate returned work, and it can increase model calls and token usage. The point is that a skill is a reusable piece of knowledge, while a subagent is a separate worker, and confusing the two means paying coordination costs for expertise that could simply have been written down. m:…0762
Meta shows what happens when the same idea runs across a whole fleet of models. The company introduced the Ranking Engineer Agent to automate the generation and testing of machine learning techniques for its ads ranking models, and says that as of March 2026, AI agents handle over 60% of all technique propagation. Engineering productivity has increased 3 times since agents began driving execution, select model combinations saw a median 4 times reduction in training compute, and Meta estimates over $45 million in compute savings by the end of 2026. Meta aims to reduce the time for a proven technique to reach all applicable models from over a year to three months. The agents use a shared knowledge base of prior training runs and model outcomes, and techniques must show improved offline metrics and pass online testing before shipping. m:…4731
The infrastructure underneath that propagation is standardization. Meta implemented a template-driven approach known as Model Templates to standardize model design into reusable building blocks, shifting iteration from a per-model problem to a shared systems problem. The company says the templates reduced per-model engineering time by 92% over two years, improved offline model quality by 0.63% on average at neutral serving capacity, and delivered 0.86% cumulative online lift covering 70% of models. The lesson is that the reusable unit, whether called a skill, a template, or a technique in a shared knowledge base, is becoming more valuable than any single run. The open question is governance: who decides what agents may learn, who can edit a skill, and who is responsible when a learned procedure goes wrong. m:…4732
What is still unresolved
- Whether skills learned from verified traces generalize to new tasks, or only speed up work the agent has already done successfully.
- Who owns and approves a shared skill inside an organization, and how a bad or stale skill gets retired before it spreads.
- Whether the benchmark gains Zep reports hold up in independent testing on other workloads and with other teacher models.
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.