Meta's Confucius Code Agent (CCA) is an open-source AI software engineer designed for industrial-scale repositories, while the Ranking Engineer Agent (REA) applies the same framework to accelerate Meta's ads ranking ML pipeline. Together, they demonstrate Meta's thesis that agent scaffolding — not model selection — determines whether an agent can handle real engineering work.
Confucius Code Agent (CCA)
Built on the Confucius SDK (a Meta + Harvard collaboration), CCA is designed for large-scale, real-world codebases — not just benchmarks.
Three-Perspective Architecture
CCA is structured around three perspectives:
- Agent Experience (AX) — What the model sees: curated context, memory state, available tools
- User Experience (UX) — How engineers interact with the agent
- Developer Experience (DX) — How the agent itself is extended and maintained
This separation is unusual and pragmatic — most agent frameworks conflate these concerns.
Hierarchical Working Memory
CCA's standout feature is its memory system:
- Persistent note-taking across sessions — the agent can write notes about what it learned that survive between invocations
- Hierarchical organization — notes are structured, not flat, so the agent can efficiently retrieve relevant context
- Token efficiency — cross-session memory reduced token usage by ~10% while improving performance
This is a direct contrast to Stripe's Minions (stateless, no memory) and suggests a different architectural philosophy: some tasks benefit from agents that learn over time.
Benchmark Performance
CCA achieved 54.3% on SWE-Bench-Pro — state-of-the-art at time of publication. More importantly, it was designed to perform on real-world enterprise codebases, not just benchmark suites.
Ranking Engineer Agent (REA)
REA applies the Confucius framework to Meta's ads ranking team — a specialized use case that demonstrates how agent patterns translate to ML engineering.
Hibernate-and-Wake Mechanism
REA's key innovation is its hibernate-and-wake pattern:
- Launch an ML training job
- Shut down the agent to conserve compute resources
- Automatically resume when the training job completes
- Analyze results, decide next steps, launch the next experiment
This is critical for ML workflows where training runs take hours or days. A traditional agent would burn compute waiting. REA's approach is both economical and practical.
Impact
Work that previously required 2 engineers per model now takes 3 engineers across 8 models. Engineers shifted from hands-on experiment execution to strategic oversight — designing experiments and interpreting results rather than running them.
Key Architectural Insight
Meta's core thesis: "How you structure the agent around the model (the 'scaffolding') determines whether it can handle real software engineering work." The model is interchangeable; the scaffolding is the differentiator. This aligns with Stripe's "the walls matter more than the model" philosophy but adds the dimension of persistent memory.
Why It's in Assess
CCA is open source, which makes it directly usable — a significant advantage over most internal-only systems. The hierarchical memory pattern and hibernate-and-wake mechanism are genuinely novel contributions. However, CCA is still relatively new and the Meta-specific integrations (internal repos, ML pipelines) limit direct applicability. Assess the memory architecture for long-running or iterative tasks, and the hibernate-and-wake pattern for any workflow involving long-running jobs (ML training, data pipelines, CI).
Key Characteristics
| Property | Value |
|---|---|
| Company | Meta (+ Harvard) |
| Systems | Confucius Code Agent (CCA), Ranking Engineer Agent (REA) |
| Architecture | Three-perspective (AX/UX/DX) with hierarchical memory |
| SWE-Bench-Pro | 54.3% (SOTA at publication) |
| Key innovations | Persistent cross-session memory, hibernate-and-wake for ML jobs |
| REA impact | 3 engineers managing 8 models (was 2 engineers per model) |
| Open source | Yes — CCA (arXiv) |
| Sources | CCA Paper, REA - Engineering at Meta |