Agent Tooling Takes Off

A year ago, most “agents” were still demos that could book a mock flight or draft a toy email. In April 2026, the tools getting attention are the ones that handle the boring plumbing: connectors, memory, approvals, and traces that show what the agent actually did. (openai.com) That plumbing is the difference between a chatbot and a worker. A production agent needs a way to reach calendars, files, and websites, plus a log of every tool call, handoff, and guardrail check when something goes wrong. (developers.openai.com) Hermes Agent is one sign of that shift. Nous Research describes it as a self-improving agent with a built-in learning loop, and its public codebase shows a large plugin surface, memory layers, gateway components, and recent packaging work aimed at real deployment rather than a weekend prototype. (hermes-agent.nousresearch.com) (github.com) The plugin piece matters because every company has a different pile of software. Hermes documents plugins and memory providers as separate components, so developers can swap in tools and storage the way a contractor swaps drill bits instead of rebuilding the whole drill. (deepwiki.com) (github.com) Rowboat is tackling a different failure mode: agents that forget everything after each prompt. Rowboat says it turns emails and meeting notes into a local linked knowledge graph, and its open-source repository describes that graph as long-lived context the system can use to draft emails, prepare briefs, and act on work already in motion. (rowboatlabs.com) (github.com) A knowledge graph is just a map of who, what, and how things connect. Instead of storing “Sarah mentioned the Series A round” as one lost note, Rowboat links Sarah, the fundraise, the meeting, and the follow-up task so the agent can retrieve the relationship, not just the sentence. (rowboatlabs.tech) (rowboatlabs.com) The other big change is that platform vendors are standardizing the roads into these systems. OpenAI’s Responses application programming interface now bundles built-in tools like web search, file search, computer use, remote Model Context Protocol servers, and tracing, which means developers no longer have to stitch five separate subsystems together before an agent can do one real job. (openai.com) (developers.openai.com) Model Context Protocol means one agreed way for models to talk to outside tools. OpenAI’s documentation now treats remote Model Context Protocol servers as a first-class tool path, which is the software equivalent of replacing a drawer full of custom chargers with one standard port. (openai.com) (developers.openai.com) The workspace products launching around this are selling the same promise in a friendlier wrapper. Genspark said on April 8, 2026 that Workspace 4.0 added desktop access and Microsoft Office integration for its “Claw” agent, pushing agents closer to the files and apps where office work already happens. (finance.yahoo.com) Big enterprise vendors are moving in the same direction. ServiceNow said on February 26, 2026 that its Autonomous Workforce agents would execute work with governance controls and that EmployeeWorks would connect Moveworks chat and enterprise search into those workflows for nearly 200 million employees. (newsroom.servicenow.com) Put together, the new story is less “look what the model can say” and more “look what the system can safely do.” The market is rewarding tools that make agents inspectable, connected, and stateful, because those are the pieces you need before anyone lets an autonomous system touch a real inbox, a real spreadsheet, or a real approval chain. (developers.openai.com) (openai.com))