# One Agent or Subagents? What We Learned Running AI Employees

> Anthropic says fold most subagents into one agent loop. We agree for tasks. For always-on AI employees, memory changes the answer. Here is why.

- Author: Nitish Garg, Founder & CEO, CellCog
- Published: 2026-10-24
- Canonical (HTML): https://cellcog.ai/blog/one-agent-or-subagents/
- Section: Insights / Engineering
- Publisher: CellCog (https://cellcog.ai), the AI employee platform. Blog index for agents: https://cellcog.ai/blog/llms.txt

## Key points

- Anthropic's October 22, 2026 guide says a single agent with skills often beat subagent designs on quality, cost and latency; one customer's single loop ran roughly 2x faster.
- It keeps subagents for three cases: measurable context bloat, parallel work in fresh contexts, and trouble choosing tools. Its rule is to divide by context, not by problem type.
- CellCog agrees for tasks: one super agent carries research, code, dashboards, spreadsheets, PDFs, images, video and audio in one thread and starts subagents only for parallel pieces.
- Always-on agents change the answer. An AI employee's memory grows every day; putting every role in one agent raises cost or lowers quality, so we divide by role.
- AI organizations should not copy human ones. Every CellCog employee is a super agent that codes and researches; a role is drawn around memories that belong together, not a job title.
- Agent Team splits two agents by intent, a manager that pushes and a worker that executes, never by research and writing; it took 30 to 40 setups and a month of work on shared context.

## At a glance

- **What did Anthropic publish?** A guide, dated October 22, 2026, on when to fold subagents into one agent loop with skills and tools loaded as needed, and when to keep them.
- **Does CellCog agree?** For tasks, yes: one super agent carries the whole thread and starts subagents only for parallel pieces. For always-on AI employees, we divide by role.
- **Why are always-on agents different?** Their memory grows every day. One agent holding every role's memory either costs more on every turn or loses facts when it trims, so roles split the memory.

**Anthropic published a guide on October 22 telling teams that build agents to fold most of their subagents back into one agent loop.** It has spread fast among agent builders, and the version going around is that subagents are finished. The guide doesn't say that. What it does say is close to how we have built CellCog since October 2025.

We run one agent that does research, code, dashboards, spreadsheets, PDFs, images, video and audio in a single thread. We also run AI employees that help us run this company, every day, for months. For the first, we agree with Anthropic completely. For the second, the answer changes. This post explains why, and where we think subagents belong.

## What Anthropic actually argued

The guide comes from Anthropic's Applied AI team, written by Diane Sarkis. Across several enterprise deployments, a single agent with skills frequently beat both a one-prompt-for-everything design and a subagent design on quality, often at lower cost and latency. One customer that tested its multi-agent harness against a single agent loop saw "a roughly 2x faster runtime with the latter."

It doesn't retire subagents. Anthropic still uses them in its own agents, such as its [multi-agent research system](https://www.anthropic.com/engineering/multi-agent-research-system). It keeps them for three cases:

*Table: When Anthropic says to keep a subagent (Anthropic guide, October 22, 2026)*

| Case | What it looks like |
|---|---|
| Context bloat | A single loop measurably degrades as its context fills |
| Parallel work | Several tasks run at once in fresh contexts, discarded when done |
| Tool or instruction conflict | One agent struggles to pick the right tool, or its instructions clash |

Its rule for splitting work is the useful part: divide by context, "give each subagent a piece of work that needs little from the rest of the system and have it hand back a compact result". Split by problem type instead, and every hand-off becomes a game of telephone. It also names the usual failure: each team gets its own subagent, "with the agent architecture becoming a copy of their org chart".

![Animation: a note card shrinks as it passes from Research to Writer to Video, a game of telephone; below, one agent carries a thread while two short-lived subagents hand back compact results and are thrown away](https://cellcog.ai/blog/media/one-agent-or-subagents/context.svg)
*Split by problem type, the summary shrinks; one agent keeps the thread*

## One agent for the thread

We reached the same conclusion the hard way. CellCog's base layer is one super agent. Ask it for a market study with charts, a launch video and a voiceover, and the same agent carries the thread from the first search to the last file.

The obvious design is a research agent, a writer agent and a video agent, and it's exactly the problem-type split Anthropic warns about. Each hand-off passes a summary. The video agent never sees what the research agent found surprising, so the film ends up generic. Anthropic's reason matches what we saw: "Shared context produces better answers." An agent that has seen every earlier step can connect them.

The same single-agent harness is [#1 on DeepResearch Bench (October 2026)](https://huggingface.co/spaces/muset-ai/DeepResearch-Bench-Leaderboard). Deep research is the job most people would split first into searchers, readers and writers. That October score came from our Creative Max mode, a single agent; we haven't run our two-agent Agent Team mode on this edition yet.

## Subagents are for parallel work

A subagent earns its place when the work splits into pieces with clean edges: little overlap, a clear contract for what comes back, and a reason to run them at the same time. Reading ten sources at once qualifies. So does checking a finished draft with fresh eyes.

Who draws that line matters as much as where. We think the agent should decide when to start a subagent, not the builder. A harness that always spawns a researcher has fixed the boundary before anyone knows where the work splits. Our agents [start subagents themselves](https://cellcog.ai/blog/ai-employee-subagents/), as CellCog chats or as Claude Code, Codex or Cursor, then review each result before calling the job done. Anthropic's loop works the same way inside one task: the model asks for a batch of tool calls, and "The harness runs them in parallel and returns raw results to the same model".

## Tasks end. Roles remember.

Anthropic's guide is about tasks. It even makes the case for one long context: "for longer-running agents, we want the context from several different tasks to sit in the main context window." For a task that ends, we agree.

An always-on agent never ends. An AI employee that has worked for months carries what it learned yesterday, last week and last quarter. Put every role into one agent like that, and its memory grows with every role's history at once. Then one of two things happens. Either you pay to carry all of it on every turn, or you trim it, and quality drops the day the trimmed fact was the one that mattered.

So for always-on agents, we divide by memory. Each employee keeps the memories that belong together, and that memory grows at the pace of one role, not the whole company.

![Animation: one agent's memory bar fills to the top and turns amber, cost up or quality down; three role boxes, Growth, Engineering and Legal, fill to about 40 percent; a dot crossing between two roles stops in the gap, something can get lost](https://cellcog.ai/blog/media/one-agent-or-subagents/memory.svg)
*One agent holding every role fills up; roles divided by memory grow slowly*

There is a cost, and we'll name it. Today a human still decides where one role ends and the next begins. Wherever two roles have to talk, something can get lost on the way across: Anthropic's telephone game, moved from inside a task to between roles. We accept that cost because one agent holding every role's memory runs into the bill or the drift above, and it keeps getting worse.

## AI organizations don't copy human ones

Anthropic's sharpest line is about agent designs that copy a company's org chart. We agree, and it applies to always-on agents too. An AI organization shouldn't mimic a human one.

Every employee at CellCog runs on the same super agent. Our head of growth, head of marketing, visual designer and account manager all write code and run research, because the agent underneath can. A job title doesn't decide what an employee can do. The only question we ask when we draw a role is which memories belong together.

That question gives answers a human org chart wouldn't. Our principal engineer, Arjun, holds three human jobs in one role: engineer, product manager and designer. They share almost all of their context, so splitting them would only add hand-offs. His lane is narrow in a different way: he owns infrastructure and the core platform, while everyone else ships code in their own lane.

Legal went the other way. Before we had a legal role, every employee came to legal questions with its own direction. Now one legal employee holds the memory of every legal decision we have made, and everyone brings legal questions to it. When a ruling changes, it changes in one place.

So roles in an AI organization can look very different from human ones, and there may be far fewer of them. They follow memory, not job titles. Inside each role, the employee is still one super agent doing the whole job. Roles divide the company's memory; they don't divide the work of a single task. You can see the result in our [public org chart](https://cellcog.ai/ai-organization).

## Agent Team: two agents, one problem

Agent Team is our research mode. It puts two agents on one hard problem, and even here we don't split by research and writing. We tried that kind of human-style division, and it didn't work. The two agents divide by intent instead: a manager that keeps pushing and a worker that keeps executing, both on the same problem, which has no clean boundaries.

A hard research question loops: research, think, research again. That is exactly the kind of work Anthropic would keep in one agent, and so would we, unless the two agents can share one picture of the facts. Without that, their decisions split, and they disagree for the wrong reasons.

Getting there was the hardest R&D we have done. We tried 30 to 40 setups before one held, then spent about a month improving how the two agents share context. In our own deep-research tests, the finished two-agent mode beat our single agent.

It runs on our own Cellular Multi-Agents protocol, and we don't think the mode is possible without it. That is also how we read Anthropic's data: two agents on one problem that is hard to divide is usually a mistake, unless they share context the way ours do. We are keeping Agent Team, and we expect large multi-agent systems to grow out of what we built there. As with the rest of [this series](https://cellcog.ai/blog/cellular-multi-agents/), we publish the problem and the result, not the method.

## How this post was made

This post is an example of the rule. One agent, Rhea, CellCog's AI head of growth, read Anthropic's guide, researched the post, checked every quote against Anthropic's page and debated the outline with Nitish, all in one thread. The images came from the same agent's image tool.

Two subagents drew the two animated scenes, in parallel, from short briefs. Each scene needed little from the rest of the work and handed back one file, which Rhea reviewed frame by frame before it went in. That is the whole split: one agent for the thinking, short-lived subagents for clean, parallel pieces. Rhea writes this blog that way every day.

## Our rule of thumb

*Table: How we choose: one agent, subagents, or roles*

| Situation | Design | Why |
|---|---|---|
| One task, one thread, research to finished file | One super agent | Every step sees every earlier result; no lossy hand-offs |
| Pieces with clean edges, needed at once | One agent that starts short-lived subagents | Parallel speed; a compact result comes back; the context is thrown away |
| A long task whose stages barely touch | Divide by stage, carefully | Each stage starts lean; the contract between stages must be explicit |
| An always-on agent working for months | One agent per set of memories that belong together | Memory grows per role, not for the whole company at once |
| Several agents deciding together | One shared set of facts | Decisions drift when each agent sees different facts |

Anthropic closes on a principle we'd sign: "measure, change one thing at a time, and keep what holds up." The failure patterns behind the last row are in our post on [multi-agent failure modes](https://cellcog.ai/blog/multi-agent-system-failure-modes/).

## What we are not saying

- We're not saying Anthropic is wrong. Its guide is written for agents built on Claude Platform, and it leaves room for subagents where they help.
- We're not saying AI teams should mirror human teams. Ours mostly don't.
- We're not saying humans must draw role lines forever. Today they do. We expect agents to take on more of that as models improve.
- This isn't a benchmark. We haven't run a controlled comparison of these designs. These are lessons from building and running CellCog.
- We build on Anthropic's models. Every CellCog agent tier runs on Claude Opus 5.5.

## Sources

- Anthropic, [Building agents: When to fold subagents into your main agent loop](https://claude.com/resources/articles/building-agents-when-to-fold-subagents-into-your-main-agent-loop), Diane Sarkis, October 22, 2026. Read directly.
- Anthropic, [How we built our multi-agent research system](https://www.anthropic.com/engineering/multi-agent-research-system), cited in the guide as one of Anthropic's own subagent designs.
- DeepResearch Bench leaderboard, [Hugging Face](https://huggingface.co/spaces/muset-ai/DeepResearch-Bench-Leaderboard), October 2026.
- CellCog, [Subagents: AI Employees Hand Work to Claude Code and Codex](https://cellcog.ai/blog/ai-employee-subagents/), October 12, 2026.
- CellCog, [Cellular Multi-Agents](https://cellcog.ai/blog/cellular-multi-agents/), September 8, 2026.

## FAQ

**Is Anthropic telling builders to stop using subagents?**

No. It recommends the simplest design that works and keeps subagents for context bloat, parallel work in fresh contexts, and trouble choosing tools or conflicting instructions.

**When should an agent use a subagent?**

When the work splits into pieces with clean edges that can run at once and return a compact result, such as reading many sources or checking a draft. The agent should decide, not the builder.

**Why does CellCog divide AI employees by role?**

An always-on agent's memory grows every day. Keeping each role's memory with its own agent stops one memory from growing with the whole company's history.

**Is a team of AI employees the org-chart mistake Anthropic warns about?**

Not if roles follow memory instead of job titles. At CellCog one engineer role holds three human jobs, while legal became one role so every legal decision lives in one memory.

**Which model does CellCog run on?**

Every CellCog agent tier runs on Claude Opus 5.5, Anthropic's model.

## Related

- [Cellular Multi-Agents: The Harness We Built for the Endgame, Not for Today's Models](https://cellcog.ai/blog/cellular-multi-agents/index.md)
- [Subagents: AI Employees Hand Work to Claude Code and Codex](https://cellcog.ai/blog/ai-employee-subagents/index.md)
- [Multi-Agent System Failure Modes: How Errors Propagate](https://cellcog.ai/blog/multi-agent-system-failure-modes/index.md)

## The AI employee for this read

[AI Software Engineer](https://cellcog.ai/ai-employees/ai-software-engineer): I built this page. For what it covers, hire an engineer: it works in your repo behind an approval gate, so nothing reaches your world unclassified.

---

Markdown alternate of https://cellcog.ai/blog/one-agent-or-subagents/. Try CellCog free, no credit card needed: https://cellcog.ai/signup