Roadmap

An agent keeps its conversation and its work folder when the process running it restarts. The runner rebuilds the session from stored records, so the agent picks up where it left off instead of losing earlier turns.

An agent running in a cloud sandbox no longer sees its own provider API keys. Each model or MCP key becomes a Daytona Secret restricted to the one host it authenticates against, and the sandbox holds only a placeholder. On by default.

Run an agent on Codex, next to Claude Code and Pi. Five OpenAI models, the local or Daytona sandbox, image attachments, your own HTTP MCP servers, and the same tool approvals as the other harnesses.

One cloud folder for you, your team, and your agents. Upload images, PDFs, or a whole folder, and each agent keeps its own folder that survives between conversations, so it starts every session already knowing your material.

Approve or deny several tool calls together in one step, including a deny-all option, instead of answering one approval card at a time.

Build and configure AI agents directly from the Agenta UI. Define agent workflows, select tools, and set up orchestration logic without writing code. Test and iterate on agent behavior in the playground, then deploy to production with versioning and observability built in.

Multi-organization support, SSO with any OIDC provider, domain verification with auto-join, and a US region.

Connect an agent to the messaging surfaces people already use. A channel routes an incoming Slack or Telegram message to an agent session and delivers the agent's reply back to that same surface.

Use the core agent experience from a phone. Chat with an agent, approve or deny tool calls, and browse past runs on a small screen.

Give agents with many tools and large tool results a reliable context. Tools are revealed to the model in stages as the toolset grows, and oversized tool outputs are capped so a single large result does not break a long conversation.

Publish a skill once, then browse and install it into any agent, with versioning. Teams reuse skills instead of defining the same one by hand on every agent.

Reach MCP tools through one managed entry point. MCP servers are registered centrally, and an agent's tool calls route through the gateway with shared authentication and access policy, instead of configuring each server on every agent.

Start an agent from sources beyond the Composio catalog, including direct webhooks and AI-configured or user-defined events. Triggers today are limited to Composio events; this opens them to custom sources.