On October 23, 2026, Google Cloud released Agent Compute Engine at AGNTCon + MCPCon, allocating a hardware-isolated microVM to each AI agent session. According to a MarkTechPost report, no credentials are stored within the sandbox; each microVM has its own independent guest kernel, root privileges, and Hyperdisk, is checkpointed to zero active resources when idle, and is restored from a snapshot on the next invocation; the service is already used to support internal workloads such as Gemini, wit
On October 22, 2026, Google Cloud announced a research reference architecture codenamed Project Chrysalis, demonstrating how to run governed self-evolving AI agents on Gemini Agent Platform and ADK. The architecture's performance in a simulated Black Friday cache avalanche failure was also disclosed: time to first mitigation was shortened from 15–45 minutes to 1.8 seconds, the HTTP 5xx error rate dropped from 72.7%–80.0% to 0.0%, and P99 latency dropped from 519.8 ms to 18.2 ms. All autonomous operations are constrained by governance from Model Armor, Cloud KMS, Binary Authorization, OPA, and two-person review HITL. There are currently no subsequent independent development facts.
On October 22, 2026, Google Cloud announced a research reference architecture codenamed Project Chrysalis, demonstrating how to run governed self-evolving AI agents on Gemini Agent Platform and ADK. The architecture's performance in a simulated Black Friday cache avalanche failure was also disclosed: time to first mitigation was shortened from 15–45 minutes to 1.8 seconds, the HTTP 5xx error rate dropped from 72.7%–80.0% to 0.0%, and P99 latency dropped from 519.8 ms to 18.2 ms. All autonomous o
Google announced that API Gateway's streaming support has entered public preview. According to the Google Developers Blog, SSE, chunked transfer, WebSocket, and gRPC streaming endpoints can share the same gateway and policies with existing APIs, with no changes needed to API specifications. Streams can stay open for up to one hour; during the preview period it must be enabled when creating a gateway, and existing gateways cannot be converted in place. Combined with model routing, clients can use a single OpenAI-compatible endpoint to dispatch to Gemini, Claude, and OpenAI gpt-oss models based on the model field in the request and return responses as a stream. The gateway's default domain is changed to a fixed format composed of the gateway ID, project number, and region, which can be written into the configuration before creation. This support is currently in public preview.
Google announced that API Gateway's streaming support has entered public preview. According to the Google Developers Blog, SSE, chunked transfer, WebSocket, and gRPC streaming endpoints can share the same gateway and policies with existing APIs, with no changes needed to API specifications. Streams can stay open for up to one hour; during the preview period it must be enabled when creating a gateway, and existing gateways cannot be converted in place. Combined with model routing, clients can use
On October 22, 2026, Google Cloud announced a research reference architecture codenamed Project Chrysalis, demonstrating how to run governed self-evolving AI agents on Gemini Agent Platform and ADK. The architecture's performance in a simulated Black Friday cache avalanche failure was also disclosed: time to first mitigation was shortened from 15–45 minutes to 1.8 seconds, the HTTP 5xx error rate dropped from 72.7%–80.0% to 0.0%, and P99 latency dropped from 519.8 ms to 18.2 ms. All autonomous operations are constrained by governance from Model Armor, Cloud KMS, Binary Authorization, OPA, and two-person review HITL. There are currently no subsequent independent development facts.
Google announced that API Gateway's streaming support has entered public preview. According to the Google Developers Blog, SSE, chunked transfer, WebSocket, and gRPC streaming endpoints can share the same gateway and policies with existing APIs, with no changes needed to API specifications. Streams can stay open for up to one hour; during the preview period it must be enabled when creating a gateway, and existing gateways cannot be converted in place. Combined with model routing, clients can use a single OpenAI-compatible endpoint to dispatch to Gemini, Claude, and OpenAI gpt-oss models based on the model field in the request and return responses as a stream. The gateway's default domain is changed to a fixed format composed of the gateway ID, project number, and region, which can be written into the configuration before creation. This support is currently in public preview.
Google notified the author that they could try the Gemini 4 Argon model. A hands-on account published by CNET editor James Bricknell on October 22, 2026, shows that he used the model to generate, through conversation, a web app that can adjust armor dimensions; after about 10 iterations it reached KeriForge 8.2, which he then exported as an STL and printed in TPU 75A. The whole process took about 18 hours, of which printing took 7 hours. He said Gemini 4 Argon completed complex tasks noticeably faster than his usual Gemini 3.5 Flash-Lite, and that it understood his intent to use the web app offline rather than develop an iPad app. The final print's dimensions matched his forearm and wrist; the surface defects came from the TPU material being too soft and the model being printed too tall, not from the model itself. There are currently no additional independent sources or official statements.
Google DeepMind, together with Google Research, the World Resources Institute (WRI), and the International Institute for Applied Systems Analysis (IIASA), released the Forest Typology (ForTy) 2020 global forest type map. The map classifies every 10×10 meter parcel of land worldwide into six forest and land-use types, and serves as a key baseline for the European Union's deforestation regulation. The team trained the model on about 1.7 million labeled locations, using AlphaEarth Foundations embedding vectors as a key input; in validation at more than 9,000 randomly sampled points, the forest type classification accuracy was 90%, and the accuracy in distinguishing forest from non-forest was 95%. A companion year-by-year time series from 2017 to 2025 is also openly available and can be used free of charge under an open license on Google Earth Engine. There are currently no further developments.
Google and Voltus announced the construction of 1 gigawatt of virtual power plant capacity in PJM Interconnection. According to Canary Media, the two parties raised the PJM grid's virtual power plant target tenfold, from 100 megawatts set in June to 1 gigawatt, with data center developers paying users to reduce electricity use during peak demand or to provide battery capacity. The report said Voltus has secured customer commitments to supply power through 2040, far longer than the three-year advance requirement typical of PJM capacity auctions. The report also noted that PJM supplies power to about 67 million people, that its cost of meeting peak demand has risen more than eightfold in recent years, and that large loads such as data centers are expected to add 30 to 34 gigawatts of demand by the early 2030s. There are currently no subsequent developments.
Google DeepMind and Cambridge have published a paper demonstrating that a Gemini-based AI agent can autonomously discover new mathematics. The agent found and provided exact rules for 11 Busy Beaver Turing machines that had no prior analysis records. According to the paper, the agent first tested edge cases in a simulator, drafted informal proofs, and then completed the proofs in the Lean proof checker; all the relevant rules were verified through Lean. The paper states that this reflects AI moving from answering mathematical questions posed by humans toward open-ended discovery. Currently, this work is presented as a paper, and there have been no reports of further progress.
Google published a paper proposing OO-Harness, a method for evolving task-specific agent harnesses by task type. The method evolves harnesses along code-class lines: each specialized harness inherits from a parent class and changes only the parts that its task cluster needs adjusted; a router then selects a specialized harness or keeps the base result based on the base run's results. The paper reports that on AutomationBench business workflows, the method scored 77.41%, versus 61.62% for the base harness and 63.92% for the global method Meta-Harness; on unfamiliar applications, it kept the base harness and maintained a 59.18% pass rate, while Meta-Harness's global harness dropped to 44.90%. The authors recommend using the base harness as a fallback for mixed workloads and routing to a specialized harness only when a run exposes a known problem.
Google notified the author that they could try the Gemini 4 Argon model. A hands-on account published by CNET editor James Bricknell on October 22, 2026, shows that he used the model to generate, through conversation, a web app that can adjust armor dimensions; after about 10 iterations it reached KeriForge 8.2, which he then exported as an STL and printed in TPU 75A. The whole process took about 18 hours, of which printing took 7 hours. He said Gemini 4 Argon completed complex tasks noticeably faster than his usual Gemini 3.5 Flash-Lite, and that it understood his intent to use the web app offline rather than develop an iPad app. The final print's dimensions matched his forearm and wrist; the surface defects came from the TPU material being too soft and the model being printed too tall, not from the model itself. There are currently no additional independent sources or official statements.
Google notified the author that they could try the Gemini 4 Argon model. A hands-on account published by CNET editor James Bricknell on October 22, 2026, shows that he used the model to generate, through conversation, a web app that can adjust armor dimensions; after about 10 iterations it reached KeriForge 8.2, which he then exported as an STL and printed in TPU 75A. The whole process took about 18 hours, of which printing took 7 hours. He said Gemini 4 Argon completed complex tasks noticeably
Google DeepMind, together with Google Research, the World Resources Institute (WRI), and the International Institute for Applied Systems Analysis (IIASA), released the Forest Typology (ForTy) 2020 global forest type map. The map classifies every 10×10 meter parcel of land worldwide into six forest and land-use types, and serves as a key baseline for the European Union's deforestation regulation. The team trained the model on about 1.7 million labeled locations, using AlphaEarth Foundations embedding vectors as a key input; in validation at more than 9,000 randomly sampled points, the forest type classification accuracy was 90%, and the accuracy in distinguishing forest from non-forest was 95%. A companion year-by-year time series from 2017 to 2025 is also openly available and can be used free of charge under an open license on Google Earth Engine. There are currently no further developments.
Google DeepMind, together with Google Research, the World Resources Institute (WRI), and the International Institute for Applied Systems Analysis (IIASA), released the Forest Typology (ForTy) 2020 global forest type map. The map classifies every 10×10 meter parcel of land worldwide into six forest and land-use types, and serves as a key baseline for the European Union's deforestation regulation. The team trained the model on about 1.7 million labeled locations, using AlphaEarth Foundations embed
Google and Voltus announced the construction of 1 gigawatt of virtual power plant capacity in PJM Interconnection. According to Canary Media, the two parties raised the PJM grid's virtual power plant target tenfold, from 100 megawatts set in June to 1 gigawatt, with data center developers paying users to reduce electricity use during peak demand or to provide battery capacity. The report said Voltus has secured customer commitments to supply power through 2040, far longer than the three-year advance requirement typical of PJM capacity auctions. The report also noted that PJM supplies power to about 67 million people, that its cost of meeting peak demand has risen more than eightfold in recent years, and that large loads such as data centers are expected to add 30 to 34 gigawatts of demand by the early 2030s. There are currently no subsequent developments.
Google and Voltus announced the construction of 1 gigawatt of virtual power plant capacity in PJM Interconnection. According to Canary Media, the two parties raised the PJM grid's virtual power plant target tenfold, from 100 megawatts set in June to 1 gigawatt, with data center developers paying users to reduce electricity use during peak demand or to provide battery capacity. The report said Voltus has secured customer commitments to supply power through 2040, far longer than the three-year adv
Google DeepMind and Cambridge have published a paper demonstrating that a Gemini-based AI agent can autonomously discover new mathematics. The agent found and provided exact rules for 11 Busy Beaver Turing machines that had no prior analysis records. According to the paper, the agent first tested edge cases in a simulator, drafted informal proofs, and then completed the proofs in the Lean proof checker; all the relevant rules were verified through Lean. The paper states that this reflects AI moving from answering mathematical questions posed by humans toward open-ended discovery. Currently, this work is presented as a paper, and there have been no reports of further progress.
Google DeepMind and Cambridge have published a paper demonstrating that a Gemini-based AI agent can autonomously discover new mathematics. The agent found and provided exact rules for 11 Busy Beaver Turing machines that had no prior analysis records. According to the paper, the agent first tested edge cases in a simulator, drafted informal proofs, and then completed the proofs in the Lean proof checker; all the relevant rules were verified through Lean. The paper states that this reflects AI mov
On October 8, the Google Developers Blog publicly released the AQuA (Ambient Quality Agent) reference implementation. This implementation has a quality Agent run alongside production Agents, responsible for clustering failed conversations, reviewing issues, and anchoring diagnoses to deployment source code snapshots. Related coverage argues that this will not make testing disappear, but rather hands over to test teams the tasks of defining anomalies, calibrating evidence, and bearing release responsibility, and presents three validation methods: implanting known faults, blind-review controls, and root-cause replay. There are currently no facts about subsequent developments.
On October 8, the Google Developers Blog publicly released the AQuA (Ambient Quality Agent) reference implementation. This implementation has a quality Agent run alongside production Agents, responsible for clustering failed conversations, reviewing issues, and anchoring diagnoses to deployment source code snapshots. Related coverage argues that this will not make testing disappear, but rather hands over to test teams the tasks of defining anomalies, calibrating evidence, and bearing release res
On October 8, the Google Developers Blog publicly released the AQuA (Ambient Quality Agent) reference implementation, which places a quality Agent alongside the production Agent, responsible for clustering failed conversations, reviewing issues, and anchoring diagnoses to deployment source code snapshots. Related reporting notes that this type of outer-loop quality inspection may still share blind spots with the production Agent in terms of models, logs, and business definitions, and that testing teams need to shift toward designing observation surfaces, calibrating anomaly labels, maintaining evidence chains, and defining escalation rules. The reporting also presents three verification methods—injecting known faults, blind-review controls, and root-cause replay—as well as a phased adoption path for small teams. There are currently no follow-up developments.
On October 8, the Google Developers Blog publicly released the AQuA (Ambient Quality Agent) reference implementation, which places a quality Agent alongside the production Agent, responsible for clustering failed conversations, reviewing issues, and anchoring diagnoses to deployment source code snapshots. Related reporting notes that this type of outer-loop quality inspection may still share blind spots with the production Agent in terms of models, logs, and business definitions, and that testin
Google published a paper proposing OO-Harness, a method for evolving task-specific agent harnesses by task type. The method evolves harnesses along code-class lines: each specialized harness inherits from a parent class and changes only the parts that its task cluster needs adjusted; a router then selects a specialized harness or keeps the base result based on the base run's results. The paper reports that on AutomationBench business workflows, the method scored 77.41%, versus 61.62% for the base harness and 63.92% for the global method Meta-Harness; on unfamiliar applications, it kept the base harness and maintained a 59.18% pass rate, while Meta-Harness's global harness dropped to 44.90%. The authors recommend using the base harness as a fallback for mixed workloads and routing to a specialized harness only when a run exposes a known problem.
Google published a paper proposing OO-Harness, a method for evolving task-specific agent harnesses by task type. The method evolves harnesses along code-class lines: each specialized harness inherits from a parent class and changes only the parts that its task cluster needs adjusted; a router then selects a specialized harness or keeps the base result based on the base run's results. The paper reports that on AutomationBench business workflows, the method scored 77.41%, versus 61.62% for the bas
On October 8, the Google Developers Blog publicly released the AQuA (Ambient Quality Agent) reference implementation. This implementation has a quality Agent run alongside production Agents, responsible for clustering failed conversations, reviewing issues, and anchoring diagnoses to deployment source code snapshots. Related coverage argues that this will not make testing disappear, but rather hands over to test teams the tasks of defining anomalies, calibrating evidence, and bearing release responsibility, and presents three validation methods: implanting known faults, blind-review controls, and root-cause replay. There are currently no facts about subsequent developments.
On October 8, the Google Developers Blog publicly released the AQuA (Ambient Quality Agent) reference implementation, which places a quality Agent alongside the production Agent, responsible for clustering failed conversations, reviewing issues, and anchoring diagnoses to deployment source code snapshots. Related reporting notes that this type of outer-loop quality inspection may still share blind spots with the production Agent in terms of models, logs, and business definitions, and that testing teams need to shift toward designing observation surfaces, calibrating anomaly labels, maintaining evidence chains, and defining escalation rules. The reporting also presents three verification methods—injecting known faults, blind-review controls, and root-cause replay—as well as a phased adoption path for small teams. There are currently no follow-up developments.