Anthropic's Claude Agent Engineering: Multi-Agent Workflows, Skills & Eval

Anthropic has launched claude.dev, a developer portal that systematically shares the internal methodology used to evolve Claude from a solo model into a collaborative multi-agent team. The site is organized into four pillars that map directly to the layers of agent engineering: AGENTS, SKILLS, ENGINEERING, and PLAYBOOKS.

AGENTS: Dynamic Multi-Agent Workflows in Claude Code

Instead of pre-written static harnesses, Claude Code now supports dynamic workflows : for each task, Claude composes and directs a custom multi-agent team on the fly. The article identifies three failure modes of single-context agents:

Agentic laziness — e.g., a security review of 50 items stops at 35 and declares "done".

Self-preference bias — when asked to score its own output, Claude is consistently lenient.

Goal drift — after repeated context compression in long tasks, constraints like "don't do X" silently disappear.

The solution rests on three primitives: agent() to spawn a sub-agent, parallel() for fan-out concurrency, and pipeline() for sequential chaining. Each sub-agent receives an independent, clean context window with a single goal, preventing cross-contamination.

Six composable patterns are provided: classifier routing, fan-out–synthesis, adversarial validation (each worker paired with a critic), generate–filter, tournament ranking, and loop-until-convergence. The Bun rewrite from Zig to Rust was accomplished using this orchestration.

Security is architected in: when handling untrusted input (e.g., public user feedback), a reader agent has read-only access while the actor agent sees only summaries , creating an architectural isolation layer rather than relying on prompt instructions. Combined with /loop, triage teams can run 24/7.

The trade-off is explicit: workflows consume significantly more tokens and suit complex, high-value tasks ; a deep research run may involve 22 agents and 1.1 million tokens. The principle: parallelism and specialization must earn back their coordination overhead.

SKILLS: Hundreds of Internal Skills Distilled into Nine Categories

Anthropic actively uses hundreds of skills , which fall into nine categories: library & API reference, product validation, data fetching & analysis, business process automation, scaffolding templates, code quality review, CI/CD, on-call runbooks, and infrastructure operations.

Three counter-intuitive insights:

Validation skills deliver the highest ROI — worth spending an engineer's full week to perfect.

Skill descriptions are not human-readable summaries but trigger conditions for the model — they must contain keywords that help the model decide "who handles this" at session start.

Gotchas (pitfall records) are the highest-signal content in any skill and should accumulate continuously with use.

A skill is not a single Markdown file but a folder: SKILL.md acts as a hub pointing to specific files — the entire filesystem becomes progressive-disclosure context engineering .

Distribution and governance: small teams commit skills directly to .claude/skills in the repo; at scale, an internal plugin marketplace lets teams install on demand.

ENGINEERING: Deleting 80% of System Prompts Made the Model Stronger

For Claude 5 generation models, Anthropic removed over 80% of Claude Code's system prompts with no measurable drop in coding benchmarks . Many old rules were "casts" for older models (e.g., "never write multi-line comments" because old models got them wrong); new models have judgment, and the casts became constraints. Internal transcripts even showed system prompts, skills, and user requests conflicting.

Six "old → new" inversions:

Give rules → Give judgment

Give examples → Design good interfaces (examples restrict exploration space)

All upfront → Progressive disclosure

Repeat emphasis → Concise tool descriptions CLAUDE.md as memory → Automatic memory

Simple specs → Rich references (HTML mockups, test suites, rubrics can serve as specs)

Practical advice: keep CLAUDE.md lightweight, spend tokens on codebase-specific gotchas; bind system prompts to product context (most valuable when building custom harnesses); prefer code references over natural language — an HTML mockup conveys design better than text or a screenshot . Run /doctor in Claude Code to slim down CLAUDE.md and skills.

PLAYBOOKS: Turning "Feels Better" into Measurable Numbers

This section answers "how to prove it works." A good eval requires four properties: task distribution matches production, stronger models score higher, frontier models still have headroom, and low variance across runs .

A common pitfall is adversarial sampling : selecting cases because "today's model fails on them" measures the model's failure fingerprint, not task difficulty. Correct approach: humans must articulate why a case is hard; cases come from production traffic, bug reports, and tickets. Scorers are chosen by output space: programmatic verification for closed outputs, LLM-as-judge for open outputs, and the judge model must not be the model under test .

Hillclimbing includes anti-overfitting guardrails: split eval set into train/test; if train improves but test flatlines, declare overfit and roll back the patch ; never paste failing examples verbatim into prompts to avoid reward hacking.

Real results: on an internal support benchmark, starting from Opus 4.8 high-effort (74.4% accuracy, $4.60 per ticket), hillclimbing reduced model size, effort, and prompts to achieve 90.5% accuracy on a held-out set at roughly one-fifth the original cost . The claude-api skill itself was hillclimbed from 66% to ~88% pass rate, uncovering two evaluation bugs where the task spec was wrong.

Closing the Loop

Anthropic's methodology forms a closed loop: use workflows to team up, skills to codify experience, new context engineering to unshackle the model, and eval to verify every step . The evolution from solo Claude to collaborative agent teams stems not from a single breakthrough but from this compounding engineering discipline.

https://claude.dev/
https://x.com/ClaudeDevs/status/2105391694741119047