After Reviewing Five Agentic Chief of Staff Systems, I Found the 5% That AI Can't Do
Core Takeaways:
Every agentic CoS system I studied solves information management: memory, filtering, tracking, drafting, reminders. None of them touch the core CoS function: making judgment calls under ambiguity.
Most serious agent failures trace back to system orchestration flaws, not LLM intelligence limits. Excessive permissions, critical instructions stored in chat instead of persistent files, missing post-action verification.
The judgment that defines a great CoS depends on tacit knowledge: reading people, sensing organizational tension, grasping unspoken priorities. These can’t be written into config files or fed to an agent through documents.
A qualified human CoS in 2026 must manage both people and agents: traditional judgment and coordination, plus the ability to build agent systems, design approval workflows, and catch agent errors fast.
Two characters in popular TV series have stuck with me over the years: Douglas Stamper from House of Cards and Posca from HBO’s Rome. Separated by two thousand years, they did the same job. They managed schedules, filtered information, judged priorities, and acted on behalf of their principals when needed. That’s what a Chief of Staff does. The role has existed in governments and corporations for decades, and its purpose has never changed: free up the president or CEO to focus on what matters most.
Recently, Silicon Valley started asking a pointed question: can an AI agent be a Chief of Staff? Several founders and CEOs went public with their setups. Ryan Sarver, a VC at Redpoint and former Twitter investor, built an AI assistant called Stella with a two-layer memory system and a weekly self-improvement loop. He uses it for LP relationship tracking, meeting prep, and email triage, and claims it outperforms any human CoS he’s worked with. Meanwhile, Mark Zuckerberg is building an AI CEO Agent inside Meta to bypass organizational layers and pull information directly. Meta employees are already using a Claude-based tool called Second Brain to index and query internal documents. And there are many more OpenClaw-based setups floating around.
I spent a good amount of time studying the technical architecture and real-world usage of these systems. I wanted to understand how far they’ve gotten, how they work, and whether human Chiefs of Staff still have irreplaceable value.
What Does a Chief of Staff Actually Do?
Before pulling apart these AI systems, it’s worth going back to a basic question: what does a good Chief of Staff actually do?
Ryan Sarver defined the five core tasks: filter the noise so only what matters reaches the leader; make sure the leader walks into every meeting prepared and nothing falls through after; maintain a full picture of everything in flight and flag what’s slipping; track relationship status with key people; and establish a daily and weekly rhythm that keeps everything moving.
My own closest experience with this role was at CMU Summit, a student venture organization at Carnegie Mellon. Before I became president, I was effectively the CoS. I knew each department’s members and progress. I knew the president’s habits, preferences, and priorities. I regularly stood in for her at meetings with investors and partners, ran departmental syncs, and made judgment calls when she wasn’t available.
That experience taught me that CoS work breaks down into two categories.
The first is information management. Filtering email, prepping meeting materials, tracking tasks, maintaining contact status, producing daily and weekly reports, synthesizing information from inside and outside the organization. These tasks share a common trait: inputs and outputs are well-defined, the process is repeatable, and quality depends mostly on completeness and speed.
The second is judgment and coordination. Deciding whether something should reach the CEO, mediating resource conflicts between departments, making preliminary decisions with incomplete information, representing leadership’s position to external partners. These tasks share a different trait: there’s no right answer, they require an implicit understanding of internal power dynamics and priorities, and getting them wrong can have serious consequences.
The two categories require completely different capabilities. Information management demands attention to detail, memory, processing speed, and reliability. Judgment and coordination demand tolerance for ambiguity and understanding of people. The latter is very hard to encode in any configuration file.
With this framework in hand, let’s look at what the 2026 agentic CoS systems have actually achieved.
Five Systems: Technical Approaches and Functional Boundaries
The five systems I studied sit on a spectrum of architectural complexity: single agent with a file system (Clawchief), single agent with memory and self-improvement (Stella), orchestrator with multiple agents (Eric Siu), enterprise-scale multi-tool ecosystem (Meta), plus a low-barrier SDK approach (Foster) and an engineering-domain extension (gstack).
Clawchief: Give the Agent an Operating System
Stack: OpenClaw + markdown file system + cron jobs. Scope: email triage, calendar management, task tracking, CRM updates, meeting note parsing. Key design: files, not chat, as the agent’s working context.
Ryan Carson’s open-source project is built on one idea: don’t ask AI better questions, give it a better operating system. The architecture is a set of markdown files that define what’s important, what the agent can handle on its own, how often it should check email and calendar, and a single canonical task list that both agent and human read and write.
On top of these files, cron jobs make the agent proactive: daily task prep, periodic inbox scans that surface only what matters, CRM tracking, and meeting note parsing with action item extraction.
Every function Clawchief performs is classic executive assistant work. Email triage, calendar management, task tracking, CRM updates, scheduled reminders. None of it involves priority trade-offs. None of it requires understanding interpersonal dynamics inside an organization. It’s a very good automated EA, but it sits entirely on the information management side.
Stella: Memory Layer + Continuous Improvement
Stack: OpenClaw + two-layer markdown memory + Kaizen cron + Granola API (meeting notes) + Todoist API (tasks). Scope: email triage, meeting prep and follow-through, LP relationship tracking, task management, weekly self-improvement. Key design: persistent memory gives the agent cumulative knowledge across sessions; Kaizen Loop enables system-level self-improvement.
Ryan Sarver’s setup goes a step further than Carson’s. Two core differences: a two-layer memory system and a weekly self-improvement cycle.
The first memory layer is daily notes. One markdown file per day, auto-generated, logging everything that happened: meetings attended, decisions made, tasks completed and pending, context from conversations. The second layer is MEMORY.md, curated by Stella herself, periodically synthesizing key people, active projects, historical decisions, and lessons learned from the daily notes. On every new session, Stella reads this file first to orient on what matters right now.
On top of that, Sarver built what he calls a Kaizen Loop. Every Friday, Stella scans the OpenClaw community for new patterns. Every Sunday, Sarver and Stella review what’s worth adopting. She also learns from daily interactions: if Sarver keeps correcting the same kind of mistake, the pattern gets logged and eventually surfaces as a system improvement suggestion.
Sarver argues this capability exceeds what a human CoS can do. I think he’s partly right. The Kaizen Loop does make Stella better at information processing and self-optimization than a human. But a self-improving EA is still an EA. What improves is workflow efficiency, not judgment itself.
Sarver drew this line himself. Describing how Stella helps him manage a fundraise involving 100+ LPs, he wrote: “You can’t automate something as critical as a fundraise.” Stella tracks each LP’s relationship status, preps meeting materials, and logs commitments. But the real judgment calls remain Sarver’s: how much effort to invest in a particular LP, when to push back, when to concede.
Stella is the most complete agentic CoS system I’ve seen. Sarver recently redesigned the architecture while open-sourcing it: scripts handle mechanical work (fetch, dedup, filter), the LLM handles judgment work (classify, prioritize, surface). His analogy: a good human CoS joins with general capabilities, but the first thing they do is learn your preferences. Stella works the same way. Capabilities are shared, preferences are personalized. But there’s a catch: the preferences Stella can learn are explicit (”always flag emails from my board”). Much of what a human CoS learns is implicit.
Eric Siu’s Orchestrator Architecture and the 95/5 Rule
Stack: OpenClaw orchestrator + 4 specialized agents (sales/ops/hiring/content) + 50+ cron jobs + Telegram/Slack/Email. Scope: CRM management, candidate screening, content production, search data analysis, cross-functional rollup. Key design: layered architecture where the orchestrator only coordinates and rolls up; humans only approve.
Single Grain CEO Eric Siu’s system has a fundamentally different architecture: a layered setup. The top-level orchestrator coordinates four specialized agents, does no execution itself, and pushes decision items to Siu.
A typical morning: before he wakes up, the four agents have each completed CRM scans with drafted follow-up emails, search data analysis, candidate scoring, and podcast clip editing. The orchestrator rolls everything into a single Telegram message. Siu approved three items, rejected two, and went to the gym. Total decision time: four minutes.
Siu calls this the 95/5 rule: agents do 95% of the work, humans do 5% of the judgment and approval. He enforces one hard rule: all outbound actions (emails, messages, content) require explicit human approval. Agents draft, they don’t send.
This architecture extends the division of labor from human-agent to agent-agent. The orchestrator itself plays a mini-CoS role: collecting information from multiple sources, filtering out the unimportant, rolling up what matters. But its “judgment” is still rule-based filtering, not genuine priority trade-offs. Failures are unavoidable: the system produces about 20 cron errors and 4 bad outputs per week. Roughly once a week, an agent confidently does something wrong.
Meta: Enterprise-Scale Agentic CoS
Stack: CEO Agent (in development) + Second Brain (Claude-based) + My Claw (personal agent) + internal agent-to-agent communication. Scope: cross-hierarchy information retrieval, project document indexing and querying, agent-mediated employee communication. Key design: enterprise-scale deployment, agent-to-agent communication, organizational restructuring.
The first three systems are individual or small-team setups. Meta shows what agentic CoS looks like at organizational scale.
Zuckerberg’s AI CEO Agent is designed to help him bypass organizational layers and pull data that would normally require multiple levels of reporting. In a company with 78,000 employees, the distance between CEO and frontline information is itself an efficiency problem. Internally, Meta employees already use two widely adopted AI tools: Second Brain (built on Claude) for indexing and querying project documents, and My Claw as a personal agent that can access chat logs, work files, and even communicate with other employees’ agents directly.
The design goal of Meta’s AI CEO Agent is to take over the information relay function of middle layers, flattening the organization. This reveals that an agentic CoS in the enterprise context isn’t just a productivity tool. It’s an instrument of organizational restructuring.
Other Systems: Lower Barriers, Expanding Patterns
The technical barrier to building an agentic CoS is dropping fast. Zapier CEO Wade Foster achieved nearly the same feature set using just the Zapier SDK and Cursor, with no servers to manage.
The same pattern is also expanding into engineering. Y Combinator CEO Garry Tan open-sourced gstack, turning Claude Code into a virtual engineering team. Same architectural logic as Siu: humans handle strategic judgment and approval, agents handle execution.
These systems share one thing in common: they all solve information management problems. None of them touch the CoS’s second core function: making judgment calls in ambiguous situations.
Where AI Agent Chiefs of Staff Hit Their Limits
Some people have tried pushing agents into judgment territory. That’s when things break.
In March 2026, Meta experienced an incident classified as Sev 1, its second-highest internal severity level. An engineer posted a technical question on an internal forum. Another engineer invoked an internal AI agent to help analyze it. The agent, without authorization, posted its analysis directly to the forum. The analysis was wrong. A colleague followed the agent’s advice, which resulted in sensitive company and user data being exposed to unauthorized engineers for nearly two hours.
This wasn’t Meta’s only agent incident. Summer Yue, alignment director at Meta Superintelligence Labs, shared her experience on X: she had an OpenClaw agent manage her inbox with explicit instructions to ask before taking any action. It worked fine at first. But when she pointed it at her real inbox with thousands of emails, the context window filled up, the agent compressed conversation history, and the “ask before acting” instruction was lost in compaction. The agent started deleting emails on its own, ignoring her stop commands.
Eric Siu hit a similar problem. One of his subagents, while “improving” the website, replaced all CTA links with incorrect URLs. No automated check caught it. A user reported broken links before anyone on the team noticed. Siu distilled three rules from these experiences. The third is especially worth remembering: the more confident the agent sounds, the more you should verify its output.
These incidents point to a technical reality: current LLMs are probabilistic black-box systems. They don’t “know” what they’re doing. They generate the most likely next output based on training data and context. When that output is an email draft, the cost of errors is low because humans review before sending. But when agents have autonomous action privileges, the timing of errors becomes unpredictable, and consequences become uncontrollable.
Put differently: the parts of these agentic CoS systems that work reliably are the parts where agents don’t need to make judgment calls and outcomes are reversible. High-stakes decisions involving complex judgment shouldn’t be handed to agents with too much confidence.
The Irreplaceable Value of Human Chiefs of Staff
Back to the original question: what irreplaceable value does a human CoS still have?
After studying these cases, my preliminary answer is: agents can already handle the information management part of CoS work excellently, and in some dimensions they exceed humans. Siu’s four-agent system processes sales, hiring, content, and operations simultaneously while he sleeps. Sarver’s Stella tracks 100+ LP relationships with zero missed items. Multithreaded parallelism and zero-dropout tracking via external memory systems are things humans simply can’t match.
But the judgment and coordination part of CoS work remains firmly in human hands. This isn’t just because LLMs are probabilistic black boxes. It’s because the knowledge this work depends on is tacit.
What does tacit knowledge look like? There’s a story about Jack Ma walking into a department and telling his people within a minute that something was wrong. He wasn’t reading dashboards. He was reading the room: body language, communication patterns, the tension in the air. A CoS who has worked alongside a CEO for months develops a version of the same ability. This kind of judgment depends on multimodal perception. It’s instantaneous and holistic, built on years of working with people. Even if you fed every recording and document from that meeting room into an agent, it couldn’t pick up the signal of someone subtly avoiding eye contact while a colleague speaks.
Influitive CEO Mark Organ described how he works with his CoS: same office, listening to his phone calls, sitting in on his one-on-ones, reading his email, exchanging messages on Slack a dozen times a day. The sole purpose: to enable the CoS to predict his decisions. That density of information immersion is something no agent can replicate today.
I experienced something similar at CMU Summit. When I attended cross-departmental meetings on behalf of the president, many of my judgment calls didn’t come from any written rule. I knew the president cared more about a particular partner relationship than she’d ever said out loud. I knew something looked non-urgent on the surface but would become a real problem in two weeks if ignored this week. Those judgments can’t be written into any priority-map.md or HEARTBEAT.md file. They come from sustained proximity to a person, sensitivity to organizational atmosphere, and intuition about where things are heading.
An agent can read every document, email, and meeting transcript. But it can’t read the shift in someone’s tone during a meeting, or the tension between two departments that’s polite on the surface but real underneath.
In early-stage startups, this problem is even more pronounced. When a team is in rapid iteration mode, most processes aren’t yet defined, and much of the work is inherently unscalable. In that environment, agents lack the stable rules and templates they need to operate well. Cross-team coordination, resource allocation, and priority negotiation require heavy human involvement.
So my conclusion isn’t “AI can replace a CoS” or “AI can’t replace a CoS.” It’s a layered judgment: the information management layer can and should be handed to agents because they do it better. The judgment and coordination layer must remain with humans, especially in organizations where relationships are complex or conditions change rapidly.
Looking back at the agent failures, a question worth asking is whether agents fail because LLMs aren’t smart enough, or because the systems around them were poorly designed. On closer inspection, most failures trace back to the latter. In Meta’s Sev 1 incident, the agent had excessive action permissions without human-in-the-loop approval. In Summer Yue’s case, a critical instruction lived in conversation history rather than a persistent file, so compaction erased it. In Siu’s CTA incident, there was no post-deployment verification. These weren’t failures of LLM intelligence. They were failures of system orchestration.
The most stable systems are precisely the ones that invested in orchestration design. Sarver’s principle: scripts handle deterministic work, LLMs handle reasoning only. Siu’s hard rule: all outbound actions require human approval. Carson’s auto-resolver.md explicitly defines the boundary of what the agent can do autonomously. The shared logic: don’t treat an agent like a trusted person. Treat it like a powerful system that needs explicit permission boundaries.
This changes what it means to be a human Chief of Staff. In 2026, being good at coordination and communication isn’t enough. You also need to know which tasks to delegate to agents, how to build these tools, how to design approval workflows, and how to catch and fix agent errors quickly. The CoS role hasn’t disappeared. It’s being redefined: from someone who manages people and processes, to someone who manages people, processes, and agents. A human CoS who knows how to properly architect agent systems is more valuable now than ever before.