I ran 100 AI agents on one laptop. The hierarchy is what made the work usable

One model doing everything runs out of room before it runs out of ability

A single session that writes the code and then checks it reads far more than it produces. It opens whole files to find one function and runs a whole test suite to see one failure. All of that stays in the same context that holds the plan and the decisions made two hours ago. By the end of a long session the model is reasoning about a transcript of the project. It rarely looks like a context problem. The typical failure looks like finished work: a change made in one file and forgotten in the file that calls it, or a test reported as passing without being run. None of it comes back as an error, which is what makes it expensive. The correction is structural: move the reading out of the room where the decisions are made. Claude Code's subagent documentation names this directly: delegating an operation that produces a lot of output keeps that output in the subagent's own context, and only the summary returns. The lead then spends its room on judgment, which is the one thing it cannot delegate.

Delegation only works when the boundary is a permission

A subagent is defined in a file, .claude/agents/<name>.md for one project or ~/.claude/agents/<name>.md for every project. Four lines of frontmatter do the work: name identifies it, description tells the lead when to delegate to it, tools lists what it is allowed to touch, and model sets which model runs it. The body is its system prompt. The tools line is the part people skip, and it is the reason a hundred agents can be trusted with a project. A researcher declared with tools: Read, Grep, Glob, WebFetch can read anything and write nothing. A reviewer given Bash can run the test command and still not edit the file it is judging. This is a permission, so it holds when the model is wrong and when the instruction is ambiguous. An instruction that says please do not edit anything holds only while the model agrees with it. The lead keeps what cannot be split: the plan, the decisions between two acceptable approaches, the integration of results that arrive separately, and the final approval. Everything else is a delegation with a boundary, and Claude Code enforces that boundary itself, because each subagent runs with independent permissions.

Fable 5.1 leads because the expensive thinking happens once

The model line accepts fable, opus, sonnet, haiku, a full model ID, or inherit to follow the main conversation. Claude Code resolves it in a fixed order: a model passed at invocation, then the frontmatter, then CLAUDE_CODE_SUBAGENT_MODEL, then the model of the main conversation. So routing lives in the definition file, and you never have to repeat it in a prompt. Anthropic's model overview is the basis for how I route. It recommends Claude Opus 5 for most workloads and Claude Fable 5.1 for demanding reasoning and long-horizon agentic work, which is exactly the shape of a lead that plans a large goal and then holds it across hours. It also lists the prices: 10 dollars per million input tokens for Fable 5.1, 5 for Opus 5, 2 for Sonnet 5 and 1 for Haiku 4.5, with Haiku on a 200 thousand token context window while the other three carry 1 million. That spread is the argument for a hierarchy rather than a fleet of equals. Planning happens once and its errors spread through every task after it, so it belongs to the most capable model available. Searching a repository, comparing two sources, summarizing a log and applying a repetitive edit are bounded jobs with a clear acceptance check, and they are the volume of the run. The documentation names cost control through faster, cheaper models as one of the reasons subagents exist. The cheap model still has to pass its check.

The task graph decides how many agents run

Claude Code sets a default ceiling of 20 subagents running at once, and spawning the twenty first fails with a concurrent subagent limit reached error. An agent may spawn agents of its own up to 3 layers below the main conversation. Both are configurable, through CLAUDE_CODE_MAX_CONCURRENT_SUBAGENTS and CLAUDE_CODE_MAX_SUBAGENT_SPAWN_DEPTH. My run reached about 100 agents in total, on one laptop, for one goal. What decides the count is the task graph. Two tasks that touch the same file stay in sequence, because parallel edits to one file produce repair work that costs more than the time saved. A task that needs an earlier decision waits for it. Parallel work is reserved for branches that can reach a finished result on their own, which is the same rule whether the wave holds 3 agents or 15. There is one more limit worth knowing before the roster grows: Claude Code warns at startup once the combined description fields pass 15,000 tokens, so keep each description to the one sentence that says when to delegate. The rule that made the difference is the last one. A report is not a completion. A task closes when the command named in the plan passes and its evidence is recorded, and until then it stays open however confident the agent sounded. That approval gate is the lead's job: Fable 5.1 reads every result and runs the checks itself before a task is marked done.

شغّلت ⁨100 agent⁩ على لابتوب واحد، وما أنجح التجربة هو ترتيب من يدير من

على لابتوب واحد، وضعت قرابة ⁨100 agent⁩ أمام هدف واحد. في الأعلى ⁨Claude Fable 5.1⁩، وأسمّيه الرأس الكبير: يدير سير العمل ويوافق على التغييرات قبل أن تدخل المشروع. وتحته ⁨agents⁩ من ⁨Claude Opus⁩ و⁨Claude Sonnet⁩ تعمل بالتوازي، لكل واحد منها قطعة محددة من العمل.

إن وصلت إلى هنا من المقطع الذي يسأل هل تستطيع فعلًا أن تدير شركة برمجيات كاملة من لابتوبك، فجوابي من التجربة أن ذلك ممكن بشرطين. الأول هرم واضح: الرأس يمسك الخطة والقرار الأخير، وكل ⁨agent⁩ تحته يأخذ مهمة لها بداية ونهاية. والثاني أن تكون حدود كل ⁨agent⁩ صلاحيات مكتوبة في ملفه، فالـ⁨agent⁩ الذي لم تعطه أداة كتابة لن يعدّل ملفًا مهما طلبت منه.

أما لماذا لا يكفي نموذج واحد يفعل كل شيء، فلأن المكان يضيق عليه حين يكتب الكود ويفحصه في الجلسة نفسها، مهما كان قويًا. ومع طول الجلسة يبدأ بتسليم عمل يبدو كاملًا، وفيه ثغرات لا تظهر في أي رسالة خطأ. توزيع العمل على ⁨agents⁩ يُبقي جلسة الرأس خفيفة، لأن كل واحد منها يحمل تفاصيله في مساحته ويعيد إليه الخلاصة.

والهرم يخفف الفاتورة أيضًا، لأن أغلى نموذج يعمل على الخطة والموافقة الأخيرة، وأغلب الـ⁨tokens⁩ تذهب إلى نماذج أرخص منه. وعمل النموذج الأرخص يمر على الفحص نفسه قبل أن يُقبل.

وعدد الـ⁨agents⁩ يحدده شكل المهام. قد يحتاج مشروع صغير إلى ثلاثة فقط، وقد تتفرع المهام في مشروع آخر إلى العشرات، والمهم أن يعرف الرأس ما الذي يعمل الآن وما الذي ينتظر غيره. وكل مهمة تبقى مفتوحة إلى أن يمر فحصها، مهما بدا تقرير الـ⁨agent⁩ مقنعًا.

الرقم والهرم من تجربتي. أما الحدود التقنية والأسعار فمن توثيق ⁨Anthropic⁩، وستجد الرابط عند كل معلومة منها.

تابعني على ⁨Instagram⁩:

التسجيل في الدورة: https://halaqa.app/enrollment?course=start-with-ai

الخطوات

Prompt

# AGENT HIERARCHY FOR ONE LAPTOP
# Use this when one goal is too large for a single session to hold.
# Lead: Fable 5.1. Implementation: Opus 5. Reading and volume: Sonnet 5.

# --- FILES ---
# CLAUDE.md
# The rules every session must see:
# - the goal and the architecture
# - the build, test and lint commands
# - what may never be touched
# - where the memory and the task file live

# .claude/agents/researcher.md
# ---
# name: researcher
# description: Reads the codebase and external sources, returns findings only
# tools: Read, Grep, Glob, WebFetch
# model: sonnet
# ---
# Report what you found with file paths and line numbers. Change nothing.

# .claude/agents/reviewer.md
# ---
# name: reviewer
# description: Checks a finished change against its acceptance command
# tools: Read, Grep, Glob, Bash
# model: opus
# ---
# Run the command named in the task. Report pass or fail with the output.

# An agent that must not write is given no tool that writes. Control comes
# from that permission, whatever the prompt says.

# --- THE LEAD PROMPT ---
# Read CLAUDE.md, then the task file. You are the lead. You keep the
# decisions, the integration and the final approval.
#
# Write the work as a checklist. Every task carries:
# 1. one bounded objective;
# 2. the files or system it owns;
# 3. the agent and the model that will run it;
# 4. the command that proves it worked;
# 5. the evidence it must return.
#
# Delegate only tasks that are independent. Two agents never edit the same
# file in the same wave. A task that needs an earlier decision waits.

# --- THE WAVE RULE ---
# Run 3 to 5 agents, then stop and reconcile: read the reports, run the
# checks yourself, resolve the conflicts, update the checklist.
#
# A report is not a completion. A task closes when its command passes and
# its evidence is recorded. When a check fails, open the smallest repair
# task that can fix it.
#
# Launch another wave only while independent work remains. Stop at the goal
# or at a stop condition you named in advance, then write the receipt: what
# shipped, what was verified, what failed, and the next exact action.