AI agents that act on your behalf inside business software are no longer a distant promise—they are being deployed now, and they are delivering hard numbers. Internal data from monday.com shows that AI coding agents have increased per-engineer pull-request throughput by more than 50% in parts of its organization. The same week, Box introduced new safety controls to contain the risks of autonomous agents, signaling that the race to deploy digital colleagues is running headlong into the need for enterprise security.
These announcements, made in mid-July 2026, illustrate the two-sided reality facing every Windows user and IT leader: AI can drastically reduce the administrative drag on knowledge work, but it also inherits the permissions of the systems it touches. The challenge is no longer whether to adopt agentic AI, but how to do it without handing over the keys to the kingdom.
What Actually Changed
Over a two-week period in July, a cluster of product launches and research reports sketched the new contours of the AI-augmented workplace.
monday.com shared operational data from its internal use of AI agents, built on Amazon Bedrock and orchestrated through a system called Sphera. Each agent has a persistent identity, a defined scope, a human manager, and a performance profile. Developers use agents for code generation, pull-request summaries, test creation, and documentation. The result: a better-than-50% increase in pull-request throughput where reusable agents are active. The company was careful to note that this is an internal metric, not an industry benchmark, and that faster code output does not automatically mean better software.
Box launched a set of AI agent controls that directly address the security concerns cited by most IT leaders as the top barrier to adoption. The features include safety boundaries that restrict what an agent can do based on content sensitivity, prompt-injection detection, activity oversight with audit trails, and mandatory human approvals for high-risk actions. This is not a theoretical framework—it is a shipping product response to the realization that an agent capable of reading email and updating files is effectively a privileged user.
Collabora released CODE 26.04, the development edition of its productivity suite, with AI capabilities embedded across word processor, spreadsheet, and presentation applications. Organizations can choose from OpenAI, Mistral AI, or self-hosted models, giving them control over where document content is processed. The word processor also achieved BITV 2.0 certification, strengthening its accessibility credentials.
Miro demonstrated its Sidekicks concept for collaborative workspaces. These AI agents pull context from tools such as Slack, Jira, Confluence, and GitHub into a shared visual board, then help teams move from vague discussions to concrete plans, diagrams, and action items—always with a human approval step before major output is created.
Maxworker, a new coordination platform, launched with the promise of automating the scheduling, CRM updating, and project tracking that, according to its market analysis, can consume up to 37 percent of a manager’s time. It links more than 25 business applications and requires manual approval for any critical action.
AmdoSoft introduced b4, a platform that merges traditional rule-based automation with AI agents that can interpret unstructured data—emails, scanned documents, invoices, contracts—and then hand off structured outputs to deterministic workflows.
A whitepaper from the Adecco Group added workforce context, noting that between 2022 and 2025 roughly 1.9 million new AI-related jobs were created, but only about 20 percent of European companies have deeply integrated AI. The report argues that the main managerial challenge is coordinating hybrid teams of people, AI agents, and eventual physical AI systems.
Finally, Google and Samsung announced that Gemini 2.0 will power over 40 applications, and they teased XR smartglasses due in autumn 2026, aiming to make AI interaction hands-free in professional settings.
What It Means for You
This wave of agentic AI is different from the chatbot era because it directly operates inside the tools you already use every day on Windows.
For everyday knowledge workers, the immediate impact will be a new type of assistant inside Microsoft 365 alternatives like Collabora, and inside collaboration hubs like Miro. Instead of asking a chatbot to summarize a document and then manually transferring the result into a project plan, you’ll be able to let an agent read the document, pull related items from other apps, draft a plan, and wait for your approval before updating calendars and tickets. The practical benefit is less time spent as the human glue between applications. The risk is that the agent may misunderstand context or act on stale information, so you’ll need to verify its outputs as carefully as you would a new colleague’s.
For power users and developers on Windows, the coding productivity gains reported by monday.com are not unique. GitHub Copilot and similar tools have already demonstrated that AI can accelerate boilerplate work. What is new is the agentic pattern: an AI with a stable identity that can open pull requests, modify configuration files, and respond to review comments. This shifts the developer role from writing every line to reviewing AI-generated changes, testing them, and ensuring that the logic is sound. The pitfall is that faster code production can mean faster bug production. Teams will need to invest in rigorous testing, code review, and deployment controls to match the new pace.
For IT administrators, the announcements from Box and Collabora highlight the governance decisions that are now unavoidable. You must decide:
- Which AI models can process your organization’s data.
- Whether to use cloud-hosted AI or self-hosted models that keep data on prem.
- What permissions each agent receives (least privilege is non-negotiable).
- How to detect and block prompt-injection attacks that try to hijack an agent’s behavior.
- How to log every agent action for audit and forensic purposes.
- Where to insert mandatory human approval for sensitive operations.
The Windows-focused implication is clear: as Microsoft continues to weave Copilot into Windows, Office, and Azure, the settings and policies available in Microsoft 365 and Intune will become the control plane for many of these agents. But many organizations also use third-party tools like Box, Miro, or custom automation platforms. Your security posture must cover them all, and it must treat each AI connector as a privileged integration point.
How We Got Here
The path from chatbot to autonomous agent has been short but steep.
Only two years ago, most workplace AI was a text box that you asked questions. It could draft an email or summarize a PDF, but you had to carry the result into your actual workflow manually. The breaking point came when vendors started giving AI access to application programming interfaces (APIs) and organizational context.
Suddenly, an AI could not only read your calendar but also check project deadlines, see attendee availability, draft an updated agenda, and—if permitted—reschedule the meeting itself. That change moved AI from a sidekick that gave advice to a participant that could take action.
The business case was clear: knowledge workers spend an inordinate amount of time on coordination. The Maxworker estimate that managers lose 37 percent of their time to administrative tasks may be bullish, but no one disputes that moving information between disconnected systems is a major productivity tax. AI agents promise to pay that tax on behalf of the human workforce, freeing people for higher-level judgment.
The technology leap was enabled by three trends:
1. Improved language models that can parse unstructured data and generate coherent plans.
2. Widespread API availability for enterprise tools—Outlook, Teams, Jira, Salesforce, GitHub, and dozens of others all offer programmatic access.
3. The normalization of AI assistants through products like Microsoft Copilot, which gave users a mental model for working alongside AI.
But the security model did not keep pace. Chatbots had limited blast radius; agents that can read email and write to shared drives inherit the risk profile of a user account. The Box announcement this month signals that the industry is now catching up, retrofitting guardrails to a flight that has already taken off.
What to Do Now
Whether you are a professional trying to work smarter or an IT leader responsible for your organization’s safety, the to-do list is concrete.
Choose a narrow starting point. Don’t try to deploy an agent that has broad access across all your systems. Pick one repetitive, high-friction workflow with clear inputs and outputs—like triaging support tickets, updating project status reports, or drafting meeting briefs from multiple sources.
Map the permissions exactly. Before connecting any AI, list what it needs to read, what it may change, and what must remain completely off limits. Grant read-only access wherever possible, and never allow an agent to delete, send email externally, commit financial transactions, or deploy code without a human sign-off.
Define success in terms that matter. Throughput metrics like “more pull requests” can be misleading. Track instead: incident rates, security vulnerability counts, mean time to recovery, change failure rate, and the percentage of AI output that requires substantial rework. For non-developer workflows, measure cycle time, error rate, employee satisfaction, and the time humans spend reviewing AI suggestions.
Build approval gates that actually work. A vague “Approve” button creates rubber-stamping. Design approval screens that show the reviewer exactly what the agent intends to do, what data and systems it will touch, why it thinks the action is appropriate, and whether the action is reversible. Tier your approvals: low-risk tasks can run automatically, moderate-risk tasks get a quick check, high-impact actions require a manager’s explicit confirmation.
Test for adversarial inputs. Assume that any document, email, ticket, or webpage the agent reads could contain hidden instructions designed to manipulate it. Box’s prompt-injection detection points to the solution: external guards that validate instructions outside the model itself. Red-team your agents before putting them into production.
Log everything. For every agent action, record the agent’s identity, the human who authorized its scope, the model and tools used, the data accessed, the action taken, and the approving person. This is your audit trail for compliance, incident response, and performance evaluation.
Train people to challenge AI. The most dangerous belief is that AI output is authoritative. Encourage employees to verify, question, and escalate. A healthy culture treats the agent as a capable but fallible junior colleague—not as an oracle.
Watch for regulation. As agents handle more sensitive work, expect regulators to demand transparency, accountability, and human oversight. The EU AI Act is already shaping requirements; similar frameworks are likely to spread.
Outlook
The announcements this month are a snapshot of a workplace that is reinventing itself in real time. Over the next year, you can expect agentic features to appear in every major collaboration and productivity suite. The smartglasses teased by Google and Samsung may seem like a novelty now, but hands-free, head-up AI interaction could soon become standard in field operations, healthcare, and logistics.
The defining question will not be what AI can do—the capabilities are already astonishing—but whether organizations can install the governance scaffolding quickly enough. The most successful ones will treat agents not as magic, not as a threat, but as powerful new workers that need careful job design, limited authority, and accountable human leadership. The office will gain a digital colleague. It is up to us to set the boundaries.