AI Agents News — Week of October 29, 2026

Copilot cloud sessions and Automations: persistent agent sessions, cross-device continuity, and self-improving automations

What changed: GitHub rolled out a connected Copilot cloud experience so cloud agent sessions persist across web, mobile, app, CLI and IDEs starting Oct. 28, 2026. Copilot Automations can now run in the cloud and learn from outcomes, surfacing suggested improvements that teams can approve or reject.

Why it matters: Engineers and operators can let agent work continue off-host (useful for long-running refactors or batch tasks) while keeping human checkpoints; teams can convert repetitive workflows into cloud automations that self-improve but still require explicit approvals to change behavior.

Try/watch: Enable cloud sessions in a staging project, require human approvals for any automation that modifies production-facing code, and log automation suggestions to measure how often the “self-improvements” reduce manual work versus introducing noisy or risky changes.

Snowflake adds explicit reasoning controls, extra system instructions, and stage mounts for its Cortex coding agents

What changed: Snowflake’s documentation shows a Oct. 28, 2026 update to Cortex Agents’ coding API that adds three public controls: a thinking_effort param to tune orchestration-model reasoning effort, instructions.system to append system-level directions, and stage_mounts to mount versioned Snowflake stages into the agent’s sandboxed runtime.

Why it matters: For teams using data-aware coding agents, these controls let operators balance speed vs. depth of reasoning, provide safer system-level constraints without editing the core sandbox, and give agents direct, versioned access to notebooks, apps, or datasets inside the environment they already use — reducing context-mismatch and fragile file-copy flows.

Try/watch: Use thinking_effort to run the same integration task at two levels (e.g., medium vs high) and compare correctness, latency, and token usage; mount a read-only stage with test data to validate reproducible agent runs before giving agents write access.

Akamai launches Akamai Inference and an Agent Sandbox (beta)

What changed: Akamai introduced two beta products — Akamai Inference, a managed model-serving service, and Akamai Agent Sandbox, an isolated runtime for executing agent-generated code and workflows — positioning its global edge cloud to host models and run agents with default outbound blocks and isolation controls.

Why it matters: If you build or buy agents that must touch sensitive systems or customer data, running models close to your users and data (edge inference) plus an execution sandbox that blocks unsafe outbound actions reduces latency and the blast radius of accidental or malicious agent behavior. That gives smaller teams a practical path to ship agentic features while preserving containment.

Try/watch: Evaluate Akamai’s sandbox using a red-team checklist: verify outbound restrictions, credential vaulting, and the ability to require human approval for high-risk actions; measure latency and cost for your real workflows against a central-cloud deployment. Monitor how the beta handles persistent agent state and long-running tasks.

Oracle and Google bring Google TPUs to Oracle Cloud (announced at Oracle AI World)

What changed: Oracle and Google announced that Google’s Tensor Processing Units (TPUs) will be offered on Oracle Cloud Infrastructure (OCI) starting in 2027, explicitly calling out support for agentic, long-context, and multi-step reasoning workloads that many production agents require.

Why it matters: For builders and buyers of enterprise agents this widens the choice of purpose-built AI hardware (an alternative to GPU-only clouds) and can materially change economics for large, low-latency agent deployments or for training larger in-house models used as private agent brains. That can cut inference cost-per-action and make persistent, long-context agents more affordable.

Try/watch: If your roadmap includes private or fine-tuned models for agents, add TPU-based pricing and performance into your total cost-of-ownership comparisons this quarter and run a proof-of-concept on representative agent workloads to see whether TPU economics beat your current GPU baseline. Track availability windows and region support through Q1–Q2 2027.

GitHub Agentic Workflows hardens agent platform with six releases

What changed: GitHub published a weekly update on Oct 26 describing six releases (v0.93.0–v0.93.5) that tightened static analysis, added safer handling for network/credential usage in integration jobs, improved failure diagnostics, and made Pi CLI 1.4.0 the default.

Why it matters: If you run CI/CD or automated issue workflows that rely on GitHub's agent tooling, these fixes reduce surprise failures and the risk that an agent exposes or misuses credentials — which makes it safer to move agent automation from experiments into staging.

Try/watch: Update a non-production repo to the v0.93.x line, run a representative agent workflow that exercises credentialed integrations, and verify the new failure logs and preflight credential checks before rolling to production.

Bubble debuts Agent 2 and MCP beta in founder AMA — editor preview and import flows

What changed: In an Oct 26 AMA, Bubble showed Agent 2 (a rebuilt Bubble AI Agent), a Bubble MCP (beta) integration that runs inside Claude and ChatGPT, in-editor previews with the Agent pinned, and an import path for apps built by other AI tools; the Agent now also checks its own work using an issue checker and screenshots.

Why it matters: For non‑engineering founders and ops teams, this lowers the barrier to iterate on app ideas created by coding agents: you can import an AI-generated app, preview and adjust it inside the visual editor, and have the Agent self-audit changes so QA overhead falls. The MCP connector also means you can use your preferred chat/code model to drive edits directly in Bubble.

Try/watch: Try the in-editor preview and the import flow with a small prototype; validate the Agent’s self-checks by forcing a known error and seeing whether the issue checker catches it. Monitor workload usage closely — Bubble flagged potential cost growth as apps scale.

Agent Tree: terminal-first observability for Claude Code agents

What changed: An open-source tool called Agent Tree launched (announced Oct 27) as a terminal-based live monitor for Claude Code agent workflows, showing task trees, decision branches, and live updates so developers can watch agent actions in real time.

Why it matters: Observability is the practical blocker for deploying coding agents safely. A terminal monitor lets engineers spot bad decisions, runaway API calls, or expensive loops early, which reduces risk when agents are permitted to change repositories or run tests automatically.

Try/watch: Add Agent Tree to a staging run of your Claude Code agents to observe decision branches, then create a short checklist for when an agent’s action must be manually approved. Watch for limits on integration depth or any data sent to third-party telemetry during monitoring.

SK Telecom treats AI agents as a managed workforce

What changed: SK Telecom published its first AI Governance Report and said it is considering AI agents as a new form of labor, including potentially linking agent performance to organizational key performance indicators. The carrier assigns agents employee ID numbers, records each agent’s role and supervisor, and monitors operating status, token use, and work history.

The company also says agents with greater autonomy or business impact must have a senior human supervisor, such as a team leader or executive. Agents can be suspended or retired when performance reviews show they are no longer needed.

Why it matters: This is a practical operating model for companies moving beyond pilots: inventory every agent, assign clear ownership, log its activity, and establish retirement criteria. Buyers should ask vendors whether their systems support equivalent audit trails and approval controls rather than relying on informal human oversight.

Try/watch: Create an internal register for every production agent with an owner, permitted tools, escalation path, activity logs, and a documented shutdown procedure.

Dell expands its AI Factory for on-premises agent workloads

What changed: Dell announced an expanded AI Factory with NVIDIA at NVIDIA GTC Berlin, including new rack-scale systems, an air-cooled two-socket server, a workstation with NVIDIA’s GB10 platform, and validated software and deployment templates for agentic workloads. Dell said more than 6,500 enterprises are using the platform as customers move AI projects from experimentation toward production.

The offering is designed to run agent workloads across deskside systems, data centers, and on-premises environments, giving enterprises an alternative to sending every task to a public cloud.

Why it matters: For regulated or latency-sensitive operators, the decision is shifting from “Can we build an agent?” to “Where should it run, and how do we control its data and tool access?” The new reference systems may shorten infrastructure work, but buyers still need workload-level cost and security comparisons against hosted services.

Try/watch: Pilot one bounded workflow on the proposed hardware and measure total cost per completed task, latency, energy use, model quality, and the number of human interventions before committing to a larger deployment.

NVIDIA signals a recurring software charge for agent infrastructure

What changed: NVIDIA CEO Jensen Huang discussed a possible annual software license of about $2,000 per GPU for software used to run AI agents. NVIDIA has not said which products would be covered or when the fee would begin, so this is a pricing signal rather than a confirmed commercial launch.

The proposal would expand NVIDIA’s business beyond selling chips and systems into recurring software revenue. It comes as enterprises seek security and management layers that constrain what agents can access while they execute tasks.

Why it matters: Infrastructure budgets for agentic AI may include a substantial software charge tied to GPU capacity, not just model API usage or hardware depreciation. Founders and IT buyers should avoid assuming that running more agents will have a linear cost: licensing, isolation, monitoring, and governance may become material components.

Try/watch: Request written pricing and entitlement details before sizing a deployment, and model the economics under both steady utilization and peak GPU capacity; monitor whether competitors offer comparable agent-runtime controls without per-GPU licensing.

The consumer SMS agent moment: a guide to the agents you can text

What changed: TechCrunch published a long rundown of consumer-facing AI agents that live in messaging apps and SMS — cataloging providers that can run multi-step tasks, keep memory, connect to email/calendars, and act proactively on your behalf.

Why it matters: Founders and small businesses who need low-friction customer touchpoints can prototype agent-driven services without building full apps — agents reachable by text remove distribution friction and let companies automate scheduling, follow-ups, reservations, and simple commerce.

Try/watch: Pilot one SMS-based agent for a single, revenue-linked workflow (booking, returns, or lead qualification) and instrument acceptance, auth friction, and complaint rates before widening use.

Dessa: Block’s local, full‑duplex voice model that moves voice agents off the cloud

What changed: Block released details and demos of Dessa, a speech-to-speech model that runs locally (consumer‑grade GPUs) and produces more natural cadence, emotion, and pauses — a capability pitched for on‑device voice agents and conversational experiences.

Why it matters: If you sell voice experiences (retail kiosks, phone-based support, or in-store assistants), a local full‑duplex model can cut runtime costs, reduce cloud‑data exposure, and improve latency — but it also raises the bar for social‑engineering risks because the agent’s voice can sound convincingly human.

Try/watch: Run a constrained on‑device pilot where the agent can only surface information or hand off to a human; measure task completion, user trust, and any increase in fraud or social‑engineering attempts.

Four major agent projects move under vendor-neutral governance

What changed: The Agentic AI Foundation added Agent Skills from Anthropic, Open Responses from OpenAI, OpenEnv from Hugging Face, and OpenSandbox from Alibaba. The projects cover reusable agent instructions, a common model interface, training and evaluation environments, and isolated execution; the foundation now hosts 10 projects.

Why it matters: Builders may be able to reduce provider-specific integration work and reuse more of their agent stack across vendors. The additions also signal that agent infrastructure is separating into practical layers: capabilities, model communication, evaluation, and safe execution.

Try/watch: Evaluate Open Responses and Agent Skills in a small prototype, but verify which features each model provider actually supports before committing to portability claims.

Daytona raises $60 million at a $1 billion valuation for long-running agents

What changed: Daytona Platforms announced a $60 million Series B led by FirstMark Growth at a $1 billion valuation, bringing total funding to $91 million. The company is building cloud infrastructure for agents that create many short-lived, stateful jobs, with sandboxed Linux, Windows, and macOS environments plus GPU access.

Why it matters: Agent workloads can require a different operating model from conventional web services: many isolated workspaces, persistent task state, tool access, and bursty execution. The funding validates demand for specialized infrastructure, while also raising questions about whether general-purpose clouds will offer equivalent primitives.

Try/watch: When estimating agent costs, model workspace creation, persistence, sandboxing, and idle time—not just model-token usage.

Google Cloud expands perimeter controls for autonomous agents

What changed: Google Cloud announced generally available improvements to VPC Service Controls, including folder-based perimeter membership, more granular exceptions, and guided troubleshooting for access denials. Google positioned the changes as a way to scale autonomous agents while keeping data inside enterprise security boundaries, with no additional charge.

Why it matters: Agents frequently request data across services and projects, making conventional network rules difficult to manage. Centralized, inherited boundaries can let operators expand agent access without giving every agent broad credentials or unrestricted movement across cloud resources.

Try/watch: Map each agent’s required data paths and test denied requests before production; monitor whether convenience-driven exceptions gradually weaken the perimeter.

Cisco introduces a benchmark for multi-step cybersecurity agents

What changed: Cisco introduced Agentic FAITH, an extension of its FAITH evaluation framework for tasks in which an agent chooses tools, interprets results, and acts through multiple steps. Its first integrated benchmark, VulnLoc, tests whether agents can locate vulnerable files in a code repository while avoiding false positives on patched code.

Why it matters: A final answer alone is not enough to judge an agent that investigates systems. Buyers and security teams need repeatable measurements for tool selection, intermediate decisions, stopping behavior, accuracy, and false alarms.

Try/watch: Add trace-level evaluation to internal agent pilots and score both the outcome and the actions taken to reach it.

TanStack AI reaches 1.0 with stable agent toolkit

What changed: TanStack published a 1.0 release of its TanStack AI stack, stabilizing 76 packages and formally supporting agents, subagents, coding-agent workflows, sandboxes, provider adapters, and an experimental harness for running agents in multiple environments. The post lists concrete additions (portable agent skills, subagents, computer-use tooling, MCP server support, and 23 provider adapters) and upgrade steps for existing 0.x users.

Why it matters: Builders who need a portable, framework-agnostic agent runtime can now rely on a stable open-source API to run chat UIs, tool-calling agents, and coding agents across providers and sandboxes — reducing lock-in and integration work when swapping model providers or embedding agents in web apps and editors.

Try/watch: If you ship agent features, run the TanStack 1.0 quickstart in a staging environment and validate your provider adapter (OpenAI, Anthropic, or local) and sandbox paths; watch the harness APIs closely — they’re marked experimental and could change in minor releases.

F5 and Google Cloud describe integration patterns to secure agent fleets

What changed: F5 published a technical blog describing how Google Cloud’s Agent Gateway can pair with F5 AI Guardrails to centralize routing, policy enforcement, and prompt/API inspection for fleets of agents at enterprise scale. The post explains an architecture where Agent Gateway handles traffic control while F5 inspects prompts and agent tool calls to block sensitive data and malicious intent in real time.

Why it matters: Enterprises moving from single agents to hundreds or thousands of agent deployments need centralized controls to avoid policy drift, accidental data leaks, and inconsistent access rules; this integration shows a practical path to inherit security and auditability without rewriting each agent.

Try/watch: Pilot Agent Gateway + F5 Guardrails on a non-production fleet, instrument event-level logging for agent-to-tool calls, and measure how many agent behaviors require human review versus automated enforcement; monitor for false positives on prompt inspection that could break legitimate workflows.

Material Security launches a Google Drive Security “agent” for workspace risk remediation

What changed: Material Security announced a Google Drive Security Agent that inspects Drive content and sharing patterns, synthesizes findings, and can remediate over-permissioned files automatically — aimed at risks introduced when third‑party AI tools get OAuth access to workspaces. The release cites an internal study about widespread exposure when agents or tools are granted access and positions the product as an automated remediation layer for that risk surface.

Why it matters: For buyers and security teams, this is a concrete example of agent-aware tooling: instead of treating agents as invisible users, security tooling now treats agent access and behavior as first-class telemetry and can automate remediation where manual triage doesn’t scale.

Try/watch: If your org allows third‑party AI integrations, inventory OAuth grants and run a scoped pilot of Material Security’s Drive agent (or equivalent) to see whether automated remediation reduces your exposure; track how often the agent’s decisions need human override to tune policies.