A Rock is a 90-day priority owned by one human seat with a defined Done state. The hardest part of running EOS® well is not setting Rocks. It is keeping Rocks alive between L10® meetings.
This is exactly the seam where agentic AI changes the math.
The classic Rock failure mode looks like this. The team sets seven Rocks at the quarterly. By Week 3, three Rocks are quietly drifting. By Week 6, two Rocks have not been touched in nine days. By Week 10, the team is renegotiating the Done state to make it achievable. The L10® Rock review surfaces the drift but cannot prevent it, because the L10® is one meeting and the drift happens between meetings.
Agentic AI does not set the Rock. The human Visionary, Integrator, and Accountability Partner still do that, the same way EOS® always asked. Agentic AI maintains the Rock between meetings.
What "agentic" actually means inside a Rock
Agentic means the AI runs on a schedule, has a defined job, talks to other systems, and escalates exceptions to a human. Not assistive. Not reactive. Active.
A Rock has natural seams for agentic involvement.
The milestone seam. Every Rock should have weekly milestones. An agent reads the milestone list and the team's task system or project system each Monday and reports green, yellow, red on each milestone with evidence. The Accountability Partner gets the report before the L10®. The Rock review becomes a confirmation, not a discovery.
The dependency seam. Many Rocks depend on something the Rock owner does not control. An agent watches the upstream system and flags the day the dependency slips. Two weeks earlier than the Accountability Partner would have noticed in the L10®.
The drift seam. A Rock that has not had a single activity logged in seven days is drifting. An agent watches the activity feed across the team's tools and flags Rock drift to the owner and the Integrator. No one has to remember to ask.
The Done state seam. The Done state is the hardest discipline in EOS® Rocks. An agent that has the Done state in its system prompt can quietly check, at the end of each week, "would a reasonable Accountability Partner agree this Rock would now be Done." If yes, the agent flags the Rock as ready for sign-off. If close, the agent surfaces what is left.
What an agentic Rock review feels like
Picture the Tuesday L10® at Week 7 of a quarter.
Without agents: the Integrator reads each Rock. The owner of each Rock says "On Track" or "Off Track." The room debates. Two Rocks are revealed to be much further behind than the owner admitted. The team adjusts. Twenty-five minutes used.
With agents: the Integrator pulls up the Rock dashboard. Each Rock has a status set by the agent that has been watching it all week, with a one-line evidence trail. The owner of each Rock confirms or contests. If the agent says Off Track and the owner says On Track, the team digs in on the discrepancy. Five minutes used. The other twenty minutes go to IDS on what actually mattered.
The agent did not change the Rock. The agent did not decide if it was On or Off Track. The agent compiled the evidence and made a recommendation. The human team still owns the call.
Why this is not "AI Rock setting"
A frequent first instinct: have AI set the Rocks at the quarterly. Have Claude or ChatGPT propose seven Rocks for the quarter based on the V/TO™ and the prior quarter's performance.
This is wrong, even though it is technically easy.
Rocks are an act of leadership and prioritization. The seven Rocks for the quarter are the seven things the leadership team is staking their credibility on. AI cannot take responsibility for that choice. The leadership team must.
What AI can do at the quarterly is brutally productive. It can summarize the prior quarter's Scorecard, surface every IDS item the team closed and every one still open, propose candidate Rocks pulled from the V/TO™ 1-Year Plan, and stress-test the proposed Rocks against the team's actual capacity. The team still picks. The team still commits. The team still owns it.
The line is clean: AI for input, humans for the choice.
What an agentic Rock SOP looks like
Every Rock should have its own SOP that the agent reads as its job description. The SOP has six fields.
- Rock title and Done state, verbatim from the quarterly.
- Accountability Partner, name and email.
- Weekly milestones, with target completion dates.
- Data sources the agent should read to assess progress.
- Escalation rule, when to flag yellow and red.
- Off-limits actions, things the agent must not do, like sending external email, creating contracts, or closing tickets.
The Integrator owns the SOP quality. The Accountability Partner owns the Rock.
Which AI to use
Either Anthropic's Claude or OpenAI's ChatGPT will run this pattern. For Rock tracking specifically, the work is mostly reading structured data, summarizing it, and producing a short status writeup. Both models do this well.
What matters more is the platform around the model. The platform needs three things: read access to your task system, scheduled execution so it runs without a human prompt, and a channel to escalate to a named human. Claude Code, OpenAI's Assistants API, and a handful of agent platforms all support this.
Pick one. Get one Rock running. Expand from there.
FAQ
Should every Rock have an agent? Eventually yes. In the first quarter of integration, pick the two or three Rocks where status visibility has historically been weakest. Add agents to those. Add to the rest at the next quarterly.
What if the Accountability Partner disagrees with the agent? The Accountability Partner wins. Always. The agent's job is to surface evidence, not to override the human. Document the disagreement and the SOP gets refined for next quarter.
Does this mean the L10® Rock review goes away? No. The Rock review gets shorter and sharper. The team still meets, the team still reviews each Rock out loud, the team still adjusts together. The agent just removes the prep tax.
Can agents set the Done state? No. Done state is a leadership decision at the quarterly. Once it is set, the agent reads it and assesses against it.
EOS®, Entrepreneurial Operating System®, Rocks™, V/TO™, Level 10 Meeting®, L10®, and Quarterly are concepts and trademarks of EOS Worldwide, LLC. This article is an independent practitioner perspective and is not affiliated with or endorsed by EOS Worldwide.