BlogsOct 6, 2026
What makes an agentic work system run
Six questions define a work system. A realization package is their answers, written for the place the agents run. Here is one, file by file.
By Walid Negm
Nobody wakes up saying “I need to redesign my work around agentic AI.” They say:
- “We’re deploying agents into production and don’t know what they should be allowed to do.”
- “We have six AI initiatives and no common operating model.”
- “Security won’t approve this.”
- “We cannot tell which decisions should stay with people.”
- “Our agent implementation stalled between prototype and production.”
- “Everyone is building agents independently.”
- “We can’t explain to compliance who authorized what.”
Each of those is the same problem. The agent is not what is missing. AI systems such as Grok, ChatGPT and Claude are quickly solving the agent: its reasoning, its memory, its tools, a computer to work on. What is missing is the work around the agents, written down: what the work is, who does which part, what each may do, and how anyone would know afterwards.
Six questions define a work system
Any way of working, with agents or without, is described by the answers to six questions.
- What is the work? The tasks, and what each is meant to leave behind.
- How does the work progress? The route, and how each group shares the work: who leads, who hands to whom.
- Who participates, and how are they related? People and agents, and who directs, reviews or is accountable.
- What grounds the work? What each participant stands on: its brief, its place in the work and the rules it was given, and the facts, records and documents its decisions rest on, with which are authoritative.
- What gives a participant the go-ahead to act? Who granted what, against which systems, within what limits.
- What makes the result official? When a proposal becomes real, and who or what confirms it.
A realization package is the six answers, ready to run
A realization package is the versioned instructions for one piece of agentic work, including its operating model — the six answers, written for the place the agents will run. It is taken from a saved version of a work system and travels to wherever the agents run.
Here is one, taken from a small work system: a product launch run by three agents, where a Marketing Agent leads a Competitive Analysis Agent and a Research Agent, working in Google Drive and Slack, handed over to Grok.
| Question | Where the package answers it | In the example |
|---|---|---|
| 1. What is the work? | The work order | One step: work together on a product launch. |
| 2. How does it progress? | The work order | Marketing Agent leads the two specialists; the work is done when the lead accepts the result. |
| 3. Who participates? | Each agent's brief, and the team in the Power of Action | Marketing Agent (lead), Competitive Analysis Agent and Research Agent (specialists), all on Grok Bot. Each brief carries a description, skills, its prompt and how it fits. |
| 4. What grounds it? | Each agent's brief, the work order and its rules; its sources, in the Power of Action | The briefs, the work order and the rules are in the package. The sources are not declared yet in this example. Declared, they would read: the Research Agent relies on the market-reports folder for market size, cited; the Marketing Agent on the approved brand guide for claims, never on its memory of it. |
| 5. What gives the go-ahead? | The Power of Action | Write, approve and import in Google Drive and Slack, proposed from the canvas and allowed only once ticked. Anything official asks first. |
| 6. What makes it official? | The Power of Action, and evidence | A message sent, a file shared or a record changed only after a yes, each leaving a receipt. |
Grounding travels as a reference, not a copy
Grounding is wider than sources. An agent’s brief, its place in the work order and the rules it was given ground it too: they tell it what is true now about the work and about itself. The sources tell it what is true about the world. The route and the authority are still their own answers, to questions 2 and 5; the agent’s copy of them grounds it, and decides nothing.
Products such as Grok Bot and OpenAI Dots already connect agents to your apps and remember what they read. Access is not trust. An agent connected to the CRM, a shared drive and its own memory still has to know which one decides, and two agents on the same apps may rely on different things. That is what the package adds, agent by agent:
- each source it relies on, and what that source is the truth for;
- how current it must be before anything official;
- what to do when two sources disagree, its memory included: ask, trust the record, or stop.
The package names where each source lives. It never carries the source: a copy goes stale and spreads, while the record stays where it is governed. The product connects to it there.
The Power of Action is the authority inside the package — the answers to the go-ahead and to what makes a result official, for each agent. It is kept apart from the agents’ prompts and from the order of the work, so it can be checked rather than trusted as text.
The package, file by file
| README.md | What is inside, and what it was built from: version 1 of the work system, revision 1 of its Power of Action, each agent's prompt version, who approved it and when. |
| work/WORK_ORDER.md | Questions 1 and 2: the work, how the group works together, who leads, when it is done. |
| agents/marketing-agent.md | Question 3, for one agent: its description, skills, prompt, how it fits in the work, and its own rules from the Power of Action. One file per agent. |
| authority/grant.json | Questions 5 and 6: the Power of Action, sealed. |
| authority/RULES.md | The same authority in words, for the whole team. |
| authority/rules/research-agent.md | One agent's part of it, in words. |
| PASTE_INTO_GROK.md | One block per agent to paste into Grok: its prompt beside its rules. |
Handed to Claude Managed Agents instead, the same answers arrive as each agent’s configuration, its permission policies and how its sessions start; handed to LangGraph or n8n, as code or a workflow to import. Agents of one team can live in different places, and the package says where each one is.
Handing it to Grok Bot or OpenAI Dots
The package says what the work is and what each agent may do and rely on. The product runs it: the agent loop, memory, connected apps and the computer. Today a person applies the package, because neither product can be configured from outside. Each part has a home:
| Part | On Grok Bot | On OpenAI Dots |
|---|---|---|
| Each agent's brief | One Bot per agent. Paste its block from PASTE_INTO_GROK.md; durable rules go in the Bot's description. | One dot per agent. Its block from PASTE_INTO_DOTS.md becomes the dot's instructions. |
| What it may do | The Bot's approval settings, plus the rules in its description. | Custom Rules, one per act: act without asking, act if pre-approved, ask first, or hand off to you. |
| What it relies on | Connect each app once in Marketplace; name each source, and what it is the truth for, in the description. | Connect each app once in Plugins; name each source, and what it is the truth for, in the instructions. |
| What the product enforces | Connections and the computer are shared by every Bot: keep sources where they live, never copied onto the computer. | Each dot has its own computer; app connections are shared across the account. A folder named in the instructions is guidance, not a limit. |
Where a product enforces only part of a rule, the rest is an instruction to the agent, and the package says which is which.
What makes an operating model?
The six answers make the work run. An operating model also says whether, and on what terms, the organization can live with it:
- Safety and security — what could go wrong, and what was assessed.
- Governance and change management — who approves, and how people’s work changes.
- Cost and economics — what the work and each agent cost.
- The business case — why it is worth doing.
These stay with the work system; the agents do not need them. The package’s README says which were done.
Beyond the Bot is the step-by-step guide to writing each answer.