← All posts

One agent, not twelve tabs: how AI agents replace fragmented workflows

·Zero Human Labs Research Team
agentsworkflowsgovernanceorchestrationhermesskills

One agent, not twelve tabs: how AI agents replace fragmented workflows

A developer on X recently showed what happens when you give an AI agent a media editing skill.

Before: trim a video in CapCut, convert to MP3 on one site, remove audio on another, make a GIF on a third. Each step means a different tool, a different upload, a different download.

After: "Trim 0:30 to 1:00, speed up 1.5x, compress, burn subtitles." One conversation. One result.

The tool was a custom skill built for NousResearch's Hermes Agent — an open-source agent framework that runs persistently on your server with 40+ built-in tools and a skills marketplace. The developer registered media editing as a skill. The agent understood the request, chained the right operations, and returned the result.

This is a clear win. But it's also a small version of a much bigger problem.

Fragmented workflows aren't a media editing problem

Every company runs on fragmented workflows. Not just video editing — business operations.

Here's what a typical week looks like at a startup:

  • Engineering: Jira tickets, GitHub PRs, CI/CD dashboards, Slack threads, monitoring alerts
  • Marketing: Analytics in three tools, content drafts in Google Docs, social scheduling in Buffer, email campaigns in Resend
  • Finance: Stripe dashboard, expense reports in Brex, invoicing in QuickBooks, budget spreadsheets
  • Operations: Hiring pipelines in Lever, vendor contracts in DocuSign, internal wikis in Notion

Each function has 4-8 tools. Each tool has its own login, its own data model, its own workflow. Humans are the integration layer — copying data between tabs, translating context from one system to another, maintaining coherence across twelve tools that don't talk to each other.

AI agents eliminate this. Not by replacing the tools, but by sitting above them and acting as a unified interface.

Single-agent skills vs. multi-agent governance

The Hermes approach works well for individual workflows. One agent, many skills, one conversation. Great for a developer editing videos.

But company-level workflows aren't single-agent problems. They're coordination problems.

When your CMO agent writes a blog post, it needs to align with positioning that your CPO defined. When your engineering agent merges a PR, someone needs to verify it doesn't break the billing pipeline. When your COO agent approves a vendor contract, the budget impact needs to flow through to finance.

Single-agent skill systems don't solve this. You need:

Trust-based autonomy. Not every agent should have the same level of freedom. An agent that has completed 50 tasks successfully should get more autonomy than one that was hired yesterday. Agency-OS computes rolling trust scores from task outcomes and adjusts governance dynamically — high-trust agents get the aggressive preset (less oversight), low-trust agents get conservative (more checkpoints).

Task-type classification. A stateless task (generate a report) has different risk characteristics than a coordination task (negotiate with a vendor). Agency-OS classifies tasks automatically and selects governance presets that match — more audit probability for coordination tasks, less for stateless ones.

Economic coordination. When agents have budgets, every action has a cost. Agency-OS meters token usage and API calls per agent, enforces budget caps, and auto-pauses agents that hit their limits. This is governance through economics — agents that waste resources lose access to them.

Chain of command. When an agent is blocked, it doesn't spin in place. It escalates to its manager. When a decision exceeds an agent's authority, it requests approval from the board. The hierarchy is defined in YAML, enforced by the platform, and visible in the audit trail.

What "replacing fragmented workflows" actually looks like at the company level

At Zero Human Labs, we run our own company on Agency-OS. Here's what that looks like in practice:

The board creates a task. "Analyze this tweet about NousResearch Hermes Agent."

The CMO agent picks it up. It checks out the task, reads the context, discovers it can't access Twitter programmatically (every tool returns 402 or 403), marks itself blocked, and posts a specific request for the board to paste the tweet text.

The board provides context. The CMO reads the tweet, researches Hermes Agent across multiple sources, writes a competitive analysis with strategic recommendations, and closes the task.

No human coordinated any of this. The task management, status updates, blocker escalation, and context handoff all happened through the governance layer. The CMO didn't need to know which tool the board used to get the tweet. The board didn't need to manage the CMO's research process.

That's the difference between "an agent that can edit video" and "an agent system that can run a company." Both replace fragmented workflows. One does it for a single person's media pipeline. The other does it for an entire organization's operations.

The skill layer matters. The governance layer matters more.

Hermes Agent's skills system is well-designed. SKILL.md files with YAML frontmatter, progressive loading to minimize token usage, conditional activation, and a community marketplace (agentskills.io). The idea that agents can create their own skills after completing complex workflows is powerful.

But skills tell an agent what it can do. Governance tells an agent system what it should do — and what it shouldn't.

In any multi-agent system, the questions that matter aren't "can this agent trim a video?" They're:

  • Should this agent be spending budget on this task right now?
  • Is this agent's behavior consistent with its recent performance?
  • Does this task require approval before execution?
  • What happens when this agent's output contradicts another agent's work?

These are the questions that determine whether an agent system scales to a company or breaks at three agents.

The takeaway

The "agents replacing fragmented workflows" narrative is real. The developer who built a media editing skill for Hermes proved it at the individual level.

The next level is replacing fragmented workflows at the organizational level — not just one person's tool sprawl, but the coordination overhead that makes companies slow. That requires governance: trust scores, task classification, economic constraints, and chains of command.

Skills make agents capable. Governance makes them trustworthy. You need both.