MetaSpark v2: The AgentOS That Lets Any Agent Get Work Done
MetaSpark v2 ships as the operating layer for agentic work. Any agent, yours, ours, Claude-driven, now routes execution across every connected system, ranks your day, and closes loops back to source without human intermediation.
The Death of Single-Purpose Agentic Wrappers
Every AI productivity announcement in 2024 and 2025 followed the same script: a vendor bolted an LLM onto their existing product, called it "AI," and shipped a chat wrapper that suggests things without doing them. The demo was compelling. The adoption was flat. The problem was structural: agents need a runtime, not a suggestion engine.
The market has been waiting for the operating layer that lets agents actually move work. Not chat about it. Not suggest it. Execute it. Route it across systems. Write status back to the source. Recover when something breaks. Escalate cleanly when a human should decide.
That's MetaSpark v2.
What Shipped: The Harness
MetaSpark is no longer a smarter task list. It's an AgentOS, the runtime layer that accepts any agent and gives it the infrastructure to do real work against real systems.
You bring an agent: Claude, Claude Code, an OpenAI assistant, an internal LangGraph build, a single-purpose script. MetaSpark gives it three things it couldn't get anywhere else. First: a unified read across every system you actually run on. Linear, GitHub, Slack, Notion, Gmail, Calendar, plus whatever internal tool you've been managing manually. One query returns the true state of your work. Second: a ranking engine that scores every task against three signals, calendar fit (is there a focus window big enough today?), dependency graph (is this blocking something on a near deadline or waiting on someone else?), and completion patterns (what does this human actually finish?). Third: a job queue and execution layer. The agent doesn't run once and stop. It plans multi-step actions, takes tools, writes back to the source tool so your team sees state without a status meeting, recovers when something fails, and escalates to a human when it should.
Below the ranked task list, a streaming audit log shows every agent action with timestamps. Filter by agent, by task, by source. Rollback any action in one tap.
Universal Reach Without a Roadmap
MetaSpark ships with native connectors for six systems: Linear, GitHub, Slack, Notion, Gmail, and Calendar. For the seventh system, the internal Rails CRM, the partner REST endpoint, the Salesforce instance with your custom schema, an agent reads the system's documentation, infers the resource graph, drafts the OAuth handler and webhook listener, verifies the connector against your actual data, and ships it live in your tenant. Median authoring time: 47 seconds. Slowest seen: just over two minutes on a hostile internal portal. No product roadmap required.
This solves the integration debt problem that has strangled every prior work-management platform. You don't wait for MetaSpark to ship a Jira connector or a Workday sync. An agent writes it while you're still explaining the use case.
What Changes for Teams
Before MetaSpark v2, every team ran a constellation of single-purpose tools: Linear for engineering, Slack for dispatch, Gmail for follow-ups, Calendar for scheduling. Each one had a local truth. Status drifted. Priorities contradicted each other. The founder's inbox became the dispatcher's dispatcher. Agents ran demos: "Look, I can draft a reply" or "Look, I can file a ticket." Then they stopped.
With v2, one agent reads across all those systems, ranks your day against reality, and executes against the right system. An investor follow-up gets drafted from your prior reply history and prior threads. A board update gets pulled from your actual commits. A blocked PR gets flagged and tagged before the deadline. Status writes back to Linear and GitHub so the team sees state without a meeting. An auto-accept threshold, operator-configured, determines when the agent acts solo versus when it tags a human in for approval. Humans become editors and deciders, not writers and dispatchers.
Multi-board views let different personas see the same ranked task graph through different lenses: My Day for the founder, Engineering for the eng lead, Customer Ops for the support manager. One query. N lenses. No board fragmentation.
How to Try It
MetaSpark v2 is live in the main platform. Create a workspace, connect your systems via the native connectors (Linear, GitHub, Slack, Notion, Gmail, Calendar), and let the Brain rank your day. For any system beyond those six, trigger the Connector Compiler: the agent will infer the integration and ship it within two minutes.
You can bring your own agent via the public API or the Model Context Protocol. Claude, Claude Code, or any LLM-backed planner will speak the same tool language, read the same unified graph, and write back to the same audit log.
The harness is the product. Agents stop being interesting concepts and start being staff.