CEO Fired Developers for AI. So They Built an Open-Source AI CEO
A CEO fired their engineering team to make room for AI. The developers got together and built an open source AI CEO to replace the executive who fired them.
The repo is called OpenExecutive. It has 2,100+ GitHub stars, hit #1 on Hacker News with 996 points and 687 comments, and it's Apache 2.0 licensed. The architecture is surprisingly serious.
Here's what it actually is, how it works, and why the HN thread turned into one of the most interesting AI discussions of the month.
What OpenExecutive Actually Is
OpenExecutive is an AI system that acts as a virtual executive team — backed by eight specialist agents, each with a domain, a model tier, and a retrieval-augmented knowledge base full of MBA-level content.
| Agent | Domain | Model | Reasoning |
|---|---|---|---|
| Chief Strategy Officer | Competitive analysis, M&A, OKRs | claude-opus-4-7 | Extended thinking (8K budget) |
| Chief Financial Officer | Financial modeling, fundraising, unit economics | claude-opus-4-7 | Extended thinking |
| Chief HR/People Officer | Hiring, comp, culture, org design | claude-sonnet-4-6 | Standard |
| General Counsel | Contracts, IP, employment law, compliance | claude-opus-4-7 | Extended thinking |
| Chief Operating Officer | Process, vendor management, scaling | claude-sonnet-4-6 | Standard |
| Chief Marketing Officer | GTM, brand, messaging, PR | claude-sonnet-4-6 | Standard |
| Chief Product Officer | Roadmap, prioritization, product strategy | claude-sonnet-4-6 | Standard |
| Board Comms Director | Board decks, investor relations, governance | claude-opus-4-7 | Extended thinking |
The user never sees the agent architecture. Every message goes to a single "Executive" orchestrator, which uses Anthropic tool use to call as many specialists as needed in parallel, then synthesizes the results into one coherent voice. No LangGraph, no CrewAI — just Python, FastAPI, and the Anthropic SDK.
From the architecture docs: "The internal agent architecture is never exposed to the user."
The Architecture That Makes It Work
Three design decisions lift this above a typical "eight prompts in a trenchcoat" setup.
1. Parallel Specialist Routing
When you ask a cross-domain question ("Should we raise a Series B?"), the orchestrator fires multiple specialists concurrently via asyncio.gather. Results come back as an ordered list matched to tool_use IDs, and duplicate specialist calls are handled correctly. The routing is model-driven — the Executive decides which specialists to consult, not hardcoded logic.
2. Two-Layer RAG with Domain Filtering
Knowledge retrieval has two layers:
Layer 1 — Built-in MBA knowledge: Curated content on strategy, finance, HR, legal, operations, marketing, and board governance, seeded into a local ChromaDB at startup. This is shipped with the app (git-tracked Markdown in
knowledge/builtin/).Layer 2 — Company-specific documents: Uploaded pitch decks, financial models, strategy docs — chunked at 512 words with 50-word overlap, embedded, and stored in ChromaDB.
Each specialist call triggers a domain-filtered semantic search. A CSO query only retrieves chunks tagged "strategy" or "marketing". A CFO query gets "finance" and "operations". This keeps irrelevant context out of the model's window.
3. Prompt Caching as a First-Class Concern
The Anthropic prompt caching API is not an afterthought here — it's designed into the rendering pipeline. The system prompt is structured in three cache-controlled blocks:
- Executive persona constant (1h TTL)
- Company profile block (1h TTL)
- Knowledge index summary (1h TTL)
RAG context and the past_decisions block are injected into the user turn, never the cached system prompt. The tool list is sorted by name before every API call. The persona prompt is a frozen constant — no f-strings, no datetime.now() in cached blocks.
Expected cache hit rate: 70–85% of input tokens after the first few turns.
4. Episodic Memory via SQLite
After every response, a background claude-haiku-4-5 pass extracts key decisions, initiatives, and advice into a SQLite database. The next session opens with a <past_decisions> block so the Executive remembers what it recommended last month.
This is the feature that made HN commenters laugh most — "This is morbidly unrealistic" — but it's a legitimate architectural pattern for agent continuity.
The Pricing Reality
OpenExecutive runs on the Anthropic Claude API. There is no local model option by default (it does support OpenRouter and local OpenAI-compatible servers as of v0.1.0, but those are opt-in).
For a startup running this as a daily advisor, you're looking at real API costs. The architecture docs are candid about this — cache misses on a busy session cost ~10x more per token. The prompt caching design exists precisely because this system would be uneconomical without it.
The HN thread had a sober take on this: "looks like a good way to run up a really large Anthropic token bill."
What the HN Thread Actually Said
The 687 comments broke into several camps, and the nuance matters more than the jokes.
1. "The revenge is the point." The top comment thread is pure schadenfreude — developers turning the tables on the executive class. Multiple comments pointed out that CEO work (strategy, relationships, decisions) is structurally the kind of thing AI is better at than engineering (where novel problems require actual creativity). The irony was not lost on anyone.
2. "This isn't a CEO. It's a management dashboard." The sharpest critique was that OpenExecutive is a reactive system — you ask it questions, it answers. A real CEO sets direction, builds relationships, manages crises, and takes responsibility for outcomes that aren't well-defined in advance. As one commenter put it: "What does a CEO do? Set vision, prioritize work, make resources available. AI can do all of that." The counterargument: "Being CEO is all about things AI structurally cannot do — inspire people, build trust, make judgment calls under uncertainty."
3. "The timing of this architecture is the real story." Multiple technical commenters noted that the quality of this project — real multi-agent orchestration, proper RAG, thoughtful caching — represents how far the open-source AI tooling ecosystem has come. Two years ago, this stack would have required a team of engineers and months of development. Today it's a weekend project for disgruntled developers with an Anthropic API key.
4. "The AI CEO will fire the developers too." The most upvoted dark-humor take: "Would be hilarious if the AI CEO also fired all the developers."
The Limitations You Should Know
OpenExecutive is not production-ready for running an actual company. Here's what it doesn't do:
No autonomous action. It answers questions. It doesn't negotiate contracts, hire people, or attend board meetings — which is the part of a CEO's job that matters.
Reactive, not proactive. The scheduler can surface follow-ups, but this isn't an agent that wakes up and decides the company needs to pivot. It waits for input.
Single-instance only. The architecture docs warn explicitly: "The scheduler claims rows via UPDATE ... RETURNING. Running two API machines would double-fire scheduled actions." This is not horizontally scalable.
Vendor lock-in. The default stack is Anthropic Claude. Prompt caching is Anthropic-specific. Switching providers would require rebuilding the caching layer.
No accountability. An AI CEO can't be fired, sued, or held responsible for bad decisions. That's fine for a toy — it's a governance problem for anything real.
The Lasting Signal
The OpenExecutive repo is a protest wrapped in a technical project. But the architecture is legitimate enough that it signals something real: the tooling for building multi-agent systems has crossed a threshold. Eight specialist agents, parallel routing, domain-filtered RAG, prompt caching, episodic memory — all in a single repo you can git clone and run with make dev.
Whether you think the CEO replacement angle is funny, naive, or dangerous, the underlying engineering quality says that multi-agent orchestration is past the prototype stage. The cost and reliability gaps are real, but the architectural patterns are converging.
And if nothing else — the developers who got fired for AI built something. That's more than the CEO who fired them probably did.
Sources: OpenExecutive GitHub repo (2.1k ★, Apache 2.0, v0.1.0 released 2026-06-30), HN discussion (996 pts, 687 comments), architecture docs, CLAUDE.md
Top comments (0)