CEO fired developers to make room for AI. Developers create open source AI CEO

TL;DR Highlight

A multi-agent open-source 'executive' fuses nine specialists into one voice with caching and memory.

Who Should Read

Backend developers designing multi-specialist agents with an orchestrator, prompt caching, and episodic memory

Core Mechanics

Many agents, one voice
- An orchestrator calls specialist agents (CSO, CFO, GC, etc.) in parallel via tool use.
- Users only see one executive voice; the internal agent structure stays hidden.
RAG context goes in the user turn only
- Each specialist retrieves built-in MBA knowledge and company docs separately from ChromaDB.
- Retrieved context is injected into the user turn, not the cached system prompt.
Prompt caching split into blocks
- The executive persona, company profile, and knowledge index are cached separately.
- No dynamic content goes in cached blocks; cache hit rate reaches up to 85% after the first few turns.
A small model extracts memory after each reply
- A background claude-haiku-5-5 pass stores decisions, initiatives, and advice in SQLite.
- The next session opens with a <past_decisions> block.
Scheduler assumes a single instance

Community Reactions

Satire aside, worth taking seriously, one commenter argued
- They said the notable part is AI set up as an organization rather than an emulated human.
- They think handling far more inputs than a human is a real advantage.
Counterpoint: LLM strategy advice follows trends
- A commenter cited an HBR piece describing 'trendslop', where LLMs converge on what's popular rather than what fits.
- It compared LLMs to a fresh MBA or junior consultant.
Prompts look generic, per commenters
- One commenter quoted lines from the repo's sales and finance prompts and found them platitude-like.
- Another who watched the demo said it looks like reactive managers, not proactive executives.
Useful for small founders, from one experience report
- A first-time founder said a similar agent acting as a 'boss' gave three tasks per day and kept memory.
- They said it helped catch things in fundraising, incorporation, and compliance.

How to Apply

Split your agent's prompt into cache blocks
- Today, split the system prompt into persona, profile, and knowledge index with separate cache points.
- Move RAG results out of the system prompt into the user turn.
Add a post-reply memory pass
- Use a small model after each reply to extract decisions and to-dos into a DB.
- Inject them at session start in a block like <past_decisions>.
Wrap specialists as parallel tool calls
- Define each domain agent as a tool and let the orchestrator call only what it needs, concurrently.
- Have the orchestrator synthesize the final answer in one tone.
Prevent double-firing in your scheduler
- Claim due jobs with UPDATE … RETURNING in your job table.
- Verify this claim logic before adding more instances.

Try It

copy & run

# Paste into Claude or ChatGPT (simplified version of OpenExecutive's structure)

You are an Executive who speaks in one coherent voice. Internally, pass through three specialist views (CFO, GC, CSO) in order, then synthesize, without revealing the internal structure.

<past_decisions>
{{past decision summary}}
</past_decisions>

<company_context>
{{company info and relevant docs}}
</company_context>

Question: {{question}}

Steps: 1) Write a one-line internal note of the key judgment from each specialist. 2) Reconcile conflicts. 3) Answer with one voice: a conclusion and three next actions. Do not output the internal notes.

Try next: after the answer, ask "Extract this conversation's decisions and to-dos as JSON" to mimic the memory pass.

Terminology

Episodic memoryA diary of decisions and advice from past conversations, retrieved in later sessions.

Prompt cachingThe server reuses the unchanged front part of a prompt to cut cost and latency. Changing the prefix breaks the cache.

OrchestratorA conductor-style agent that delegates work to specialist agents and assembles the final answer.

MCPA standard that lets AI tools use external servers' tools and data. Like a USB port for AI.

ChromaDBA vector DB you can run locally. It embeds documents and finds semantically similar chunks.