At Databricks, we use and construct brokers extensively, from coding with them at scale to delivery agent merchandise like Genie. However though the capabilities of brokers have gotten a lot better, working with them feels clunky. As customers, we frequently have 4-5 brokers open directly (coding brokers, Gemini search, and so forth) and spend our time copy-pasting textual content between them and Docs, Slack, and different collaboration instruments. And as agent builders, we’re on a treadmill to enhance our brokers by combining the most recent harnesses, SDKs and fashions. The issue is that LLM capabilities are wrapped into an agent harness, and these harnesses have completely different interfaces that make combining them or swapping them troublesome.
So we constructed Omnigent: a meta-harness that sits above the brokers you already use (Claude Code, Codex, Pi, or customized brokers) and makes them interoperable elements of a richer system. Omnigent targets the issues the place a single harness stops: it provides simple methods to compose a number of brokers, management them with superior insurance policies, and collaborate reside with teammates.
We imagine folks will quickly work with brokers by way of this new layer, the meta-harness. That’s why right now we’re open sourcing Omnigent below Apache 2.0.
Why construct a meta-harness?
We adopted coding brokers early throughout our 5,000+ member engineering workforce and constructed hundreds of brokers for patrons. That have satisfied us that the frontier of agent engineering is transferring up a degree. The perfect outcomes not come from a single mannequin in a single harness: Harvey beat a frontier mannequin on high quality and price by giving an open-source employee mannequin a frontier advisor it may possibly name, Anthropic constructed its analysis product as a lead agent orchestrating parallel subagents, and our personal Genie makes use of completely different LLMs for planning, search, and code era. Engineers are altering how they work, too: as an alternative of prompting one agent at a time, they design loops that drive complete groups of brokers.
These patterns span a number of harnesses, fashions, and other people, however every harness solely understands its personal classes. To mix brokers, govern them, and work on them with different folks, you want a layer above the harness. Omnigent is that layer, and it supplies:
- Composition. Mix a number of fashions, harnesses, and methods with out rewriting code, and change between Claude Code, Codex, Pi, and your individual brokers with one-line adjustments.
- Management. Stateful, contextual insurance policies that monitor agent actions and implement guardrails like value budgets and permissions on the meta-harness layer, not through prompts.
- Collaboration. Share reside agent classes through URL and evaluation information in them collectively, so teammates can evaluation, remark, and steer brokers collectively in actual time.
How Omnigent works
Omnigent introduces a standard interface above command-line brokers and agent SDKs to allow you to simply mix and interchange them, after which focuses on the shared issues the place a harness stops. The important thing perception is that nevertheless every agent harness calls into its LLM internally, the interface to customers is similar: messages and information in, textual content streams and power calls out. Thus we constructed a standard API that wraps each terminal-based coding brokers (Claude Code, Codex, Pi, and so forth) and SDKs (OpenAI Brokers, Claude Brokers SDK, and so forth).
On high of this interface, the present model of Omnigent provides the next key options:
- Actual-time collaboration: you possibly can invite different folks to view your agent session, touch upon information in its workspace, and even ship instructions, so your classes and dealing directories turn into the principle place you collaborate.
- A number of interfaces to the identical agent: when you join an agent reminiscent of Claude Code to the Omnigent server, you possibly can entry it on the internet, cellular, Mac OS native app, or APIs.
- Cloud execution: launch any agent by yourself machine or on hosted sandbox suppliers like Modal and Daytona, for secure collaboration in a airtight surroundings.
- Contextual safety insurance policies: Omnigent’s safety insurance policies transcend the easy “permit X / deny Y” of coding brokers, to trace dynamic state about every session and make smarter choices. For instance, you possibly can say that after an agent downloads a brand new bundle from npm, it ought to require human approval to git push, or that it ought to solely be capable to write to docs it created, not any doc.
- Value insurance policies: One of many issues we monitor dynamically is every session’s LLM value. For instance, you possibly can ask Omnigent to pause an agent and ask to proceed after each $100 it spends.
- Robust OS sandbox: In Omnigent, we embody a versatile OS sandbox from our safety workforce with the power to flexibly lock down OS entry and intercept and rework community requests (e.g., don’t let an agent ever see your GitHub safety token, however as an alternative, inject it solely within the egress proxy on authorised requests).
- Multi-harness authoring: Specify a customized agent as a YAML and port it throughout harnesses with a one-line change, or mix subagents utilizing completely different harnesses in the identical agent.
These options are simply scratching the floor of what could be achieved on the meta-harness layer, nevertheless, and we count on to see much more concepts quickly from our workforce and the open supply group. Some gadgets on our roadmap embody computerized optimization on the meta-harness degree with GEPA, code-based introspection inside brokers just like MemEx and RLM, an Omnigent Server MCP so brokers can work throughout your classes, and extra harnesses. We’ve additionally made Omnigent simple to deploy on a variety of infrastructure, together with Fly.io, Railway, Modal and Daytona sandboxes, and plenty of LLM suppliers, and we welcome patches for extra integrations.
A brand new layer for working with brokers
Most of the greatest shifts in our trade got here from transferring to a brand new layer of abstraction: for instance, whereas engineers used to handle particular person processes and servers, they’ll now handle a complete fleet through cloud programs like Kubernetes and Terraform.
We predict brokers are on the identical level right now. Every harness is its personal silo, with its personal context, its personal controls, and its personal method of working, and none of it carries over if you change instruments. Furthermore, many issues intrinsically span harnesses, together with composition, safety and collaboration. A meta-harness lifts your work above any single harness, so your classes, insurance policies, and abilities stick with you regardless of which agent or mannequin is working. The fashions and harnesses will maintain altering as the sphere evolves; the layer you’re employed at should not need to.
We’re constructing that layer within the open, and we might love so that you can construct it with us.
Strive it out
Omnigent is open supply in alpha right now.
