Production-ready agent configurations implementing the agentic patterns from Anthropic research:
- Building Effective Agents - 7 workflow patterns from augmented LLM to autonomous agents
- Measuring Agent Autonomy in Practice - Risk-aware oversight and autonomy calibration from millions of interactions
cp -r agents/ ~/.claude/agents/
cp instincts/base-instincts.yaml ~/.claude/instincts/| Agent | Pattern | Description |
|---|---|---|
| intelligent-router | Routing + Pattern Selection | Classifies tasks, selects simplest effective pattern. 8-level complexity spectrum + risk-aware adjustment. |
| prompt-chainer | Prompt Chaining | Sequential LLM pipeline with programmatic quality gates between steps. |
| voting-selector | Parallelization (Voting) | Same task across multiple agents, selects best via comparison or consensus. |
| evaluator-optimizer | Evaluator-Optimizer | Iterative generate-evaluate-refine loop until quality threshold met. |
| Agent | Research Finding | Description |
|---|---|---|
| risk-assessor | Risk-autonomy distribution | Scores actions on risk (1-10) and autonomy (1-10). Decision matrix: GO / LOG / CONFIRM / STOP. |
| oversight-monitor | Post-deployment monitoring | Post-execution auditor for autonomous sessions. Flags scope creep, risky actions, intent drift. |
| autonomy-calibrator | Deployment overhang | Calibrates agent freedom per task. Four profiles: full autonomy to human-directed. |
| File | Description |
|---|---|
| _anthropic-workflow-patterns.md | Complete mapping of all 7 patterns to agents with decision flowchart |
| base-instincts.yaml | Behavioral rules: agent-initiated stops, risk-aware autonomy, deployment overhang |
User Request
|
v
intelligent-router
|-- [Step 0] Select simplest effective pattern
|-- [Step 0.5] Risk-aware adjustment
|-- [Step 1] Intent classification
|-- [Step 2] Route to pattern
v
+-----------+------------+-----------+-------------+----------------+
| Single | Prompt | Parallel | Orchestrator| Evaluator |
| Agent | Chainer | (Voting/ | Workers | Optimizer |
| (default) | (gates) | Section) | (dynamic) | (refine loop) |
+-----------+------------+-----------+-------------+----------------+
|
risk-assessor (gates)
autonomy-calibrator (profiles)
oversight-monitor (post-audit)
1. Start Simple - Router defaults to single agent for ~70%% of tasks. Orchestration only when needed.
2. Agent-Initiated Stops - Claude pauses for clarification more than humans interrupt it. Five categories encoded in instincts:
- Present choices (35%%) | Gather diagnostics (21%%) | Clarify request (13%%) | Request credentials (12%%) | Get approval (11%%)
3. Risk-Aware Autonomy - Low risk = full autonomy. Medium = log + monitor. High = confirm first. Critical = human-directed.
4. Deployment Overhang - Agents can handle more than they exercise. For low-risk tasks, push further rather than stopping prematurely.
5. Post-Execution Monitoring - oversight-monitor audits what agents did after autonomous sessions.
Designed for Claude Code. Standard agent format:
---
name: agent-name
description: What this agent does
tools: Read, Write, Edit, Bash, Glob, Grep
---
Agent instructions here...- Building Effective Agents - Anthropic Engineering Blog
- Measuring Agent Autonomy in Practice - Anthropic Research, Feb 2026
- Claude Agent SDK
- Anthropic Cookbook - Patterns
MIT