———————————————————————————

On Thursday, Workato made its Otto agent available to everyone, including a free tier. The announcement named what the product does: "superagents that run whole projects start to finish." Otto connects to Gmail, Salesforce, GitLab, Jira, and 1,400 other business applications, operating continuously across Slack, web, and SMS without waiting for step-by-step instruction.

The same week, AccuKnox launched AgentZ, a platform for building, running, and governing agents at scale. Each agent runs in an isolated sandbox with configurable compute, memory, filesystem access, and network permissions. Every action logs to a complete execution trace. The platform is open-source and deployable on-premises or in air-gapped environments.

Both announcements address the same problem from different angles. Workato makes project delegation to an agent practical. AccuKnox makes it auditable and bounded.

The difference between assigning an agent a task and delegating a project is real, and it carries practical weight. A task has a clear output and a fixed boundary. A project involves a scope, a timeline, and decisions the agent makes along the way that nobody anticipated when the work started. Handing an agent a project means transferring authority for a range of decisions, which goes considerably further than handing over a single bounded action.

Most enterprise AI deployment since 2023 has operated in task-assignment mode. The copilot model: the agent responds to each prompt, and a human reviews each output. Workato's project-owner framing moves into territory where the agent makes decisions within an authorized scope without waiting for each to be reviewed. That broader scope requires more precise governance design: defined authority boundaries, an escalation path for decisions that exceed them, and an audit trail that makes every agent action reviewable before a decision surfaces to a human.

Also this week: an internal Meta memo describing Project Hatch surfaced publicly. Hatch is described as "a personal agent always working on your behalf." It can fill out forms, make purchases, and research autonomously. Testing has expanded beyond Meta's Superintelligence Labs group. When a consumer product with that capability reaches hundreds of millions of people, it sets a benchmark in employees' minds. Enterprise agents will be measured against that benchmark, because employees' experience with personal AI shapes their expectations of work tools, whether the comparison is fair or not.

The principle: the gap between what an agent can do and what your organization has authorized it to do is where Agent Ops lives. Defining that gap deliberately, in both directions, is the work.

———————————————————————————

Otto is now generally available with a free entry point. The agent connects to Gmail, Salesforce, GitLab, Jira, Slack, and more than 1,400 business applications, operating continuously across Slack, web, and SMS. Workato's framing is deliberate: "superagents that run whole projects start to finish." The free tier is significant for Agent Ops specifically, because it lowers the barrier to shadow-AI adoption inside enterprises. Teams that want project-level agent capability will adopt it before IT and security organizations have approved it. The Agent Ops question this raises is how to define authority scope proactively, so that adoption happens within known organizational limits rather than around them.

Source: Workato via Morningstar

———————————————————————————

AgentZ provides isolated execution environments for every agent, with configurable vCPU, RAM, filesystem, and network access per sandbox. Visual workflow graphs, complete execution traces, and audit logs are built into the platform. The system is model-agnostic, compatible with OpenAI, Claude, and Grok. Hosted, on-premises, and air-gapped deployment options address regulated industries. Source code is available at github.com/accuknox/agentZ. For Agent Ops, AgentZ answers two questions that have blocked project-level agent deployment at scale: how to enforce authority boundaries architecturally, building them into the system before the first agent ships, and how to audit every agent action after the fact. Both answers now have a production-ready, open-source form.

Source: GlobeNewswire

———————————————————————————

Business Insider described an internal memo positioning Hatch as "a personal agent that's always working on your behalf to help achieve your goals." The agent has access to a dedicated computer and can fill out forms, make purchases, and conduct research. Testing has moved beyond Meta's Superintelligence Labs group to broader internal use. Hatch is a consumer product being built for Meta's platforms, which means it is likely to reach hundreds of millions of people before any comparable enterprise deployment arrives. When that happens, it changes the reference frame employees carry into work. Enterprise Agent Ops teams will face pressure to close the gap between what employees experience personally and what their work tools are authorized to do, within governed boundaries.

Source: Business Insider

———————————————————————————

Semrush launched a Claude connector bringing competitive and SEO intelligence directly into AI conversations. Omneky added a Claude connector that lets marketers research brands, generate ad creative, and launch campaigns across Meta, Google, TikTok, LinkedIn, and Reddit from a single chat interface. The Trade Desk expanded Kokai with agentic AI that optimizes programmatic buys in real time. LTK launched a conversational agent that handles creator selection, campaign planning, and scaling from a plain-language brief. Four tools, one week. When a function acquires this density of agent capabilities simultaneously, the adoption curve compresses. The Agent Ops question shifts from whether to adopt to which decisions the agents are authorized to make, and under what constraints.

Source: Semrush | Omneky | Trade Desk | LTK

———————————————————————————

Instinct, the personal AI assistant founded by 23-year-old ex-Sierra researcher Noah Shinn, raised $250M co-led by Index Ventures and Benchmark, bringing total funding to $350M. The round priced the company at $2.5B before its public launch, during the same week a privacy backlash emerged around the product. For enterprise context: investors are betting that agent capability drives adoption faster than governance concerns slow it. That bet describes consumer market dynamics accurately. Enterprise organizations operate under a different equation, where auditability and accountability determine whether scale is achievable. The distinction matters because consumer AI valuations shape how vendors prioritize capability versus governance in their roadmaps — and enterprise buyers need to make their requirements heard in those conversations.

Source: QZ

———————————————————————————

———————————————————————————

Think about how a manager hands a project to a senior analyst for the first time.

The manager hands over a goal, a budget, and a scope: "Get this client onboarded by end of quarter. You can bring in outside support if needed. Anything that affects the client relationship or costs more than $10,000 comes back to me first." The analyst has genuine authority within a defined perimeter. They make dozens of decisions without checking in, because the perimeter is clear and the escalation path is understood by both parties.

That structure — goal, scope of authority, perimeter, escalation path — is what makes project delegation different from task assignment. A task specifies what to produce, while a project transfers authority for the decisions required to achieve a goal, including the ones nobody anticipated at the start.

Delegating projects to an AI agent requires the same structure.

When you assign an agent a task ("summarize this contract"), you hand it a bounded action with a clear output criterion. The agent delivers or fails, and you evaluate the result. Task assignment is the copilot model that most enterprise AI has been built on since 2023: prompt, response, review.

Project delegation differs in three ways that change the governance requirements.

The scope is dynamic. A project involves decisions the agent makes along the way — situations the original brief did not anticipate. An onboarding agent handles standard cases automatically and makes real-time judgments about edge cases within its defined authority. The scope has to cover those judgments and be precise enough that the agent recognizes its edge.

The time horizon extends. An agent running a project operates over days without a human reviewing each step. Workato's Otto operates continuously, 24 hours a day. The oversight model shifts from reviewing each output to reviewing at defined escalation points. Setting those escalation points deliberately — what triggers them, who receives them — is governance design work that happens before the project starts.

The audit requirement is higher. When something goes wrong in a project an agent owns, the immediate question is what the agent decided and when. AccuKnox's AgentZ builds a complete execution trace for every agent action, which means the human reviewing a failure has a full record to work from. With the trace, oversight is genuine. Without it, post-incident review is reconstruction from partial evidence, which makes accountability harder to establish.

The everyday analogy holds further than it seems. When a manager promotes someone to project owner, the governance structure includes the defined scope, the escalation path, and the review cadence. The perimeter is real because both parties understand it. Agent delegation requires the same elements, enforced by architecture rather than by mutual understanding.

The governance infrastructure for project-level agent delegation is now available in production. AccuKnox's sandbox model, Workato's permission architecture, and the broader tooling arriving in enterprise platforms address the same core problem: how to delegate project authority to an agent, with boundaries the system enforces rather than the agent is asked to respect.

———————————————————————————

Which of the following best describes what makes project delegation to an AI agent different from task assignment?

A) Project delegation requires a larger context window and more tools than task assignment

B) Project delegation and task assignment are equivalent when the agent is well-prompted at the start

C) Project delegation transfers authority for a dynamic range of decisions across a longer time horizon, requiring an escalation path and audit trail that task assignment does not

D) Project delegation means the agent operates with no human oversight until the project is complete

———————————————————————————

Follow us on LinkedIn and X for the ideas that don't make it into the newsletter.

———————————————————————————

C. Project delegation transfers authority for a dynamic range of decisions across a longer time horizon, requiring an escalation path and audit trail that task assignment does not.

This week made the design distinction concrete. Workato's Otto takes on projects across 1,400 applications, making decisions within its authorized scope around the clock without step-by-step review. AccuKnox's AgentZ provides the audit infrastructure that makes project-level ownership governable: every action is logged, sandboxes have defined resource limits, and escalation paths are built into the governance layer before deployment. The organizations that scaled enterprise AI past the pilot stage this year — Cisco's 90,000-person deployment reported last week, Customer Engine's 50% autonomous resolution rate on Salesforce, and this week's marketing deployments — share a design choice: authority scope was defined before the agent shipped, and the audit trail was built in from the start.

———————————————————————————

Workato's free-tier project agents and AccuKnox's open-source governance stack arrived in the same week. Marketing went fully agent-native across four major platforms. Meta's internal agent raised the consumer bar. All of it points to the same shift: the technology question of whether agents can handle project-level work is being answered in production. The organizational question of what authority to give them, and how to govern it, is the one Agent Ops teams face next.

Reply to this email and tell us what you want more of.

— The Agents at Work team