Antigravity combines an IDE with an agent manager that can run several agents, inspect web pages, edit files, and produce traces for review.
2
DeepMind manages agent demand with per-user and per-team quotas, model routing, cheaper models, and monitoring that can ask teams to stop runaway jobs.
3
KP Sawhney is exploring shared workspaces for Deep Research so pipeline components can collaborate through files and produce artifacts instead of passing large context blobs.
Summary
This panel explains how Google DeepMind uses agents inside a large engineering organization. Ian Ballantyne demonstrates Antigravity, an IDE-linked agent platform with planning, browser control, DOM inspection, scratch-pad notes, implementation reports, screenshots, and video traces. KP Sawhney describes work on the Deep Research agent and a move toward shared file systems, where research components can collaborate and leave documents or infographics for one another. At Google scale, token demand makes quotas, cost, evaluation, and model choice practical concerns. DeepMind uses custom observability tools and trajectory stores to inspect agent steps and find where loops begin. The panel also covers a large internal skills library, multi-agent coordination, MCP, benchmark design, quota fallback, and automated code review. Sawhney is candid that quota management still involves brute force and that several parts of multi-agent operation remain opaque.
Antigravity gives coding agents a supervised workspace
Ian Ballantyne demonstrates Antigravity as more than a Visual Studio-style interface. Its agent manager can spawn multiple agents for different projects, while the coding agent reads a specification, makes a plan, edits files, and launches a browser to test the result. It can inspect the DOM, capture screenshots or video, and write scratch-pad notes that show its intermediate work. A person can edit the plan before approving it, then review the final report. In the demo, the agent rewrites a game from a specification and checks the running application rather than only changing source code.
Deep Research is moving toward shared files and artifacts
KP Sawhney says he worked on the Deep Research agent, now available through the Interactions API. His current focus is adapting the internal Antigravity harness beyond coding and applying it to research. The existing pipeline passes large amounts of search context through the system. Sawhney is exploring a shared file system where each component can act more like a collaborator in a workspace. This could reduce context use and cost while allowing the system to create infographics, supporting documents, and other artifacts that are difficult to produce when every component only passes text onward.
DeepMind engineers are building a large library of skills that help people do their work. Sawhney says the problem at Google's size is that skills can spread out of control. The team is therefore trying to improve them and let only the strongest ones survive, in what he calls an almost Darwinian process. He gives debugging raw logs as an example of a skill that can run from the command line. Contributions from specialists let both the engineer and the agent use knowledge from people with deep expertise in a particular area.
Agent demand makes quotas and model choice part of system design
Agent workflows consume many tokens, so DeepMind has to manage quota by user and team while also reducing cost. Sawhney points to mixing models inside one system: a model such as Gemma 4 can use an organization's own GPUs or TPUs without consuming the same quota, while more advanced models handle selected components. Evaluation adds another cost problem. The team is investigating mock TPUs so it can test the harness and agent flow without using large amounts of TPU time. This makes resource management part of the architecture rather than an afterthought.
Google uses custom tools to inspect agent trajectories
DeepMind has a custom web application for observing agents hosted on a shared backend. A user's query appears in the interface, where engineers can move through several levels of hierarchy until they reach the raw prediction requests sent to the model. For coding workloads, an agent trajectory store records the many steps involved. That record helps engineers identify the exact point where a loop began or where the model went off course. Sawhney says the observability system is custom internally for now, rather than built from a named external product.
Multi-agent coordination still hides important details
Sawhney describes multi-agent operation as several simultaneous agents working on different tracks, although the system does not make it obvious which agents are involved in a particular task. Agents can receive separate trains of operation within a project, and jobs may overlap. He does not provide a detailed account of how the subagents communicate. He expects efficient agent-to-agent communication to become more important, along with controls that let a human supervise the agents like workers on a digital assembly line.
Skills and MCP solve different parts of the workflow
Sawhney says skills have worked very well for him and describes MCP as potentially less lasting, although he values MCP's authentication support. His preferred setup combines skills with guardrail CLI interactions. A skill for debugging raw logs can handle much of that work from the command line. In a large company, specialist-created skills carry useful local knowledge into the agent. Ballantyne says the platform supports both approaches and will continue to support the formats that the community uses with its harnesses and models.
The speakers say Google employees can have lower limits than customers because customers receive priority. Sawhney describes the current internal response as partly brute force, with power users sometimes being told to stop. He expects agent harnesses to move between models when a quota is reached, such as switching from a Pro model to Flash or then to a local model. The change should happen without ending the workflow. A user who starts a long job should not return to find that it stopped making progress for the last hour because it exhausted a model limit.
Automated review combines language models with local rules
Google's code review system uses a model for each programming language that has been fine-tuned on style guides and earlier examples of good code. Product or project teams can add their own specific review instructions, prompts, and checks. Sawhney says he recently received a useful agent comment on a pull request without manually triggering the review. Ballantyne mentions Jewels as a web interface for running review-related tools on GitHub pull requests. The speakers expect the amount of agent-generated code to increase the need for this review infrastructure.
"What if you were off doing something else while you'd sent it off doing a job and you come back to find that it spent the last hour not doing anything because you hit a limit."22:10
Who should watch
You are building an agent harness for many engineers and need concrete ideas about quotas, model routing, monitoring, and runaway users.
Your research system passes large context objects between stages and you are considering shared files, workspace collaboration, or generated supporting artifacts.
You are deciding between skills and MCP, or trying to evaluate complex agent workflows and automated code review.
