Compound Engineering: Accelerating the Compound Evolution of AI Agents

AI agents often forget what they learned in previous sessions because most are stateless, causing repeated mistakes even with powerful models such as GPT‑5, Claude 4.5, and Gemini 3.0 Pro. This memory loss is identified as the primary obstacle to agent effectiveness.

Every.to proposes a solution called Compound Engineering (CE) . In a simple illustration, a researcher at Every.to opens their computer in the morning to see Claude automatically review code and generate check‑ins for three pull requests (#219, #234, #241). The agent not only remembers the past three months of experience but also proactively applies it.

“Changed variable naming to match pattern from PR #234, removed excessive test coverage per feedback on PR #219, added error handling similar to approved approach in PR #241.”

The core insight is that the real challenge for agents is how to remember previous learning and avoid repeating errors before becoming smarter . CE creates a learning loop where each piece of work becomes a teaching moment, turning every bug into a category‑level rule that future agents can reuse, much like financial compounding.

Four Core Mechanisms of Compound Engineering

1. Teach Through Work

Knowledge is captured and encoded at the moment decisions are made, making teaching part of the workflow rather than an after‑the‑fact task.

2. Category‑Level Prevention

Failures are treated as learning opportunities; fixing a bug also extracts the underlying pattern, preventing an entire class of similar issues.

3. Living Knowledge Base

The knowledge repository continuously evolves—new patterns are added, outdated ones are pruned, and contradictions are resolved, resembling biological evolution.

4. Feedback Loop Compression

Beyond faster code generation, the entire development cycle (plan → execute → review → compound) is compressed through parallel orchestration, automated verification, and rapid iteration, shrinking the time from days to minutes.

Productivity is expressed as Productivity = (Code Velocity) × (Feedback Quality) × (Iteration Frequency) . When code generation approaches instant speed, the bottleneck shifts to feedback quality and iteration speed, not to writing code.

Traditional solutions—larger context windows, Retrieval‑Augmented Generation (RAG), and fine‑tuning—are insufficient because they focus on making the model smarter rather than building a system that learns.

Frustration Detector Example

Kieran built a detector that automatically identifies user frustration, generates improvement reports, and iteratively refines tests. Claude writes a test, sees it fail, adjusts the prompt, reruns the test, and repeats until all ten attempts pass, demonstrating the “teach through work” loop.

Applying CE to Industry Scenarios

Two major hurdles appear when moving from coding to domain‑specific tasks: (1) agents lack industry expertise, and (2) the knowledge they receive is static. Anthropic’s Skills package addresses the first by packaging domain knowledge into organized folders that can be loaded on demand. CE then transforms this static knowledge into evolving expertise.

In industry, knowledge is proprietary, expertise develops over years, and the resulting moat is deep. By continuously extracting patterns from real outcomes—e.g., using a context + strategy + outcome triple—companies can build a compounding advantage that competitors cannot easily copy.

Three‑Layer Skills Evolution

Layer 1: Static Skills – agents have knowledge but no judgment.

Layer 2: Adaptive Skills – agents apply judgment by learning conditional probabilities from observed cases (e.g., IF high‑value THEN empathy‑first with 82% confidence from 45 samples).

Layer 3: Collective Skills – knowledge is shared across teams, multiplying learning speed (e.g., five lawyers each contribute 100 cases, yielding a five‑fold advantage).

Case Study: Collection Agency

Traditional collection uses fixed scripts; CE + Skills enables dynamic segmentation, outcome‑driven metrics, and strategy upgrades. Data showed that customers who responded with empathy‑first and disclosed difficulties had an 82% success rate, while aggressive pressure on unresponsive customers yielded only 67%.

The workflow evolves from static rules to judgment: after three months, additional dimensions (communication responsiveness, difficulty disclosure) are added; after six months, contextual judgment is applied; after a year, human‑in‑the‑loop refinements handle edge cases.

Case Study: HR Recruitment

Recruitment feedback loops are long (3–12 months). Analysis of 1,000 candidates revealed that side projects correlate with higher performance (8.2 vs 7.1 points). Further causal analysis showed that side projects are proxies for faster learning speed and self‑drive, not direct causes.

By grouping candidates with similar backgrounds and comparing outcomes, the team extracted a rule:

IF side‑project AND high‑learning‑speed THEN higher performance

, illustrating how CE turns raw data into actionable expertise.

Key Reflections

The fundamental issue is not model intelligence but system design: enabling each work episode to become a learning moment and converting bugs into category‑level defenses. Effective CE systems must also know when to forget outdated patterns.

When code generation becomes near‑instant, the limiting factor shifts to verification and learning. In domain‑specific settings, knowledge is proprietary and the moat deepens, making the CE + Skills combination potentially more valuable than in pure coding.

Adopting this mindset—trusting but verifying, moving from micromanaging each decision to cultivating a self‑improving system—unlocks compounding productivity gains across both software development and broader industry applications.

Ultimately, Compound Engineering combined with Skills accelerates the compound evolution of AI agents, turning every interaction into a source of lasting expertise.