AI agents are evolving beyond simple question answering to performing complex multi-step tasks for employees, generating detailed records of their decisions, approaches, tools, and outcomes. According to reports from Business Standard, researchers at Google and Microsoft are exploring ways to transform this operational data into reusable knowledge, procedures, and testing environments for future agents. Google's Dream-RSI (recursive self-improvement) research provides a clear example, using records of previous discovery processes as replayable environments where new exploration strategies can be tested against past results without requiring agents to repeat the original work. The emerging approach goes beyond conventional AI memory by capturing experience from completed tasks and making it available for agents working on later tasks, fundamentally changing how companies could treat AI-generated work history.
Google's Dream-RSI system demonstrated significant performance improvements in experimental testing. As reported by Business Standard, the system recorded discovery trees generated during online exploration, converted them into a reusable simulator pool, and used alternative exploration strategies against those records before deploying selected strategies for another round of exploration. Google reported that this approach reduced discovery-agent calls by up to 162 times against one existing system while maintaining or improving discovery quality in several tested settings. The system essentially uses the structure of previous work to evaluate different approaches for future tasks, with the immediate objective being reducing the cost of experimentation while the underlying idea has broader implications for work reusability.
Microsoft Research is addressing the challenge of managing growing interaction histories through its PlugMem research, which converts raw interactions into structured knowledge rather than simply storing old conversations. According to Business Standard, the system extracts two broad types of information: facts and reusable skills, with facts describing learned information and skills capturing successful execution patterns. Microsoft's Foundry Agent Service introduces procedural memory, analyzing agent activity to identify successful patterns, inefficient routes, and missing steps, then converting these findings into procedures describing when particular approaches should be used. Microsoft reported approximately 5% improvement in evaluations on STATE-Bench and Tau-Bench when procedural memory was enabled, with this distinction becoming important as agents take on recurring business processes where successful patterns can be reused across future cases.
Thoughtbot has developed a comprehensive library of agent skills and plugin marketplaces for AI tools like Claude, Cursor, and Copilot, offering proven approaches to building better software through reusable instructions. According to Thoughtbot, these skills package expertise so agents can apply it when needed rather than repeatedly prompting with the same context and process. Key offerings include the Rails Audit Skill that guides agents through comprehensive code audits based on thoughtbot practices, Atomic Commits for focused, incremental commit guidance, and Designer Consulting Skills that evaluate meeting notes and project updates against best practices. The company has spent more than 20 years refining practices for designing and building maintainable software, with skills available on their Skills GitHub page and skills.sh/thoughtbot for community use and improvement.
The volume of work being delegated to AI agents is experiencing significant growth across enterprise sectors, with a July 2026 study called Bad Memory testing malicious instructions across four models and finding that malicious payloads could persist and affect later tasks. As reported by Business Standard, OpenAI's August 2026 Enterprise Signals report indicates that Codex accounted for 64% of combined Codex and ChatGPT output tokens among enterprise customers as of June. The report also found that weekly active enterprise Codex users had grown sharply since February in areas beyond engineering, including legal, sales, recruiting, and marketing. This expansion creates substantial potential for accumulating machine-performed work history that could contain successful workflows, failed approaches, and established procedures, with the volume of work delegated to agents increasing significantly as organizations shift from AI assistance towards delegated execution.
The growing value of agent experience introduces new security considerations as stored memory gains influence over future behavior. According to Business Standard, the Bad Memory study tested malicious instructions across Anthropic's Claude Code and OpenAI's Codex, finding that getting agents to write malicious content into their own memory using untrusted external content was difficult, but once a malicious payload was present, it could persist and affect later tasks. Microsoft has highlighted similar risks, noting that persistent memory can allow attackers to influence agents over time rather than through single prompts, with researchers describing scenarios where malicious instructions stored during one interaction trigger actions in later, unrelated interactions. This creates a trade-off where the more valuable the memory becomes, the more critical it is to ensure the information entering that memory can be trusted, making the quality of stored experience as important as its quantity.
