agentflow

分类: 工具与效率 | 上传者: shouc | 下载: 0 | 版本: v1.0（最新）
使用 AgentFlow 构建和运行多代理管道。当用户想要并行、按顺序或迭代循环编排 codex、claude 或 kimi 代理时使用。当用户提到多代理工作流程、扇出任务、代码审查管道、迭代实施循环、在 EC2/ECS 上运行代理或需要多个 AI 代理一起协调的任何任务时触发。还可以触发“agentflow”、“pipeline”、“agent graph”、“fanout”、“shard”或“run codex on Remote”。
更新日志: Source: GitHub https://github.com/shouc/agentflow
目录结构

当前层级: tree/master/skills/agentflow/
📄 SKILL.md 7.7 KB
SKILL.md

---
name: agentflow
description: Build and run multi-agent pipelines using AgentFlow. Use when the user wants to orchestrate codex, claude, or kimi agents in parallel, in sequence, or in iterative loops. Trigger when the user mentions multi-agent workflows, fan-out tasks, code review pipelines, iterative implementation loops, running agents on EC2/ECS, or any task that needs multiple AI agents coordinated together. Also trigger for "agentflow", "pipeline", "graph of agents", "fanout", "shard", or "run codex on remote".
---

# AgentFlow

Build multi-agent pipelines where codex, claude, and kimi work together in dependency graphs with parallel fanout, iterative cycles, and remote execution.

## Quick Start

```python
from agentflow import Graph, codex, claude

with Graph("review-pipeline", concurrency=3) as g:
    plan = codex(task_id="plan", prompt="Plan the work.", tools="read_only")
    impl = claude(task_id="impl", prompt="Implement:\n{{ nodes.plan.output }}", tools="read_write")
    review = codex(task_id="review", prompt="Review:\n{{ nodes.impl.output }}")
    plan >> impl >> review

print(g.to_json())
```

Run: `agentflow run pipeline.py`

## Imports

```python
from agentflow import Graph, codex, claude, kimi       # agents
from agentflow import fanout, merge                    # parallel shards
from agentflow import shell, python_node, sync         # utility nodes
```

## Nodes

Create agent nodes with `codex()`, `claude()`, or `kimi()`. Required: `task_id`, `prompt`.

```python
codex(
    task_id="name",              # unique ID (required)
    prompt="...",                 # Jinja2 template (required)
    tools="read_only",           # "read_only" | "read_write"
    timeout_seconds=300,
    retries=1,
    success_criteria=[{"kind": "output_contains", "value": "PASS"}],
    target={...},                # execution target (local/ssh/ec2/ecs)
    env={"KEY": "val"},
)
```

## Dependencies

Use `>>` to set execution order:

```python
plan >> [impl, review]       # plan before impl AND review (parallel)
[impl, review] >> merge      # both before merge
```

## Template Variables

Prompts are Jinja2 templates rendered at runtime:

```
{{ nodes.plan.output }}              # output of completed node
{{ nodes.plan.status }}              # "completed", "failed"
{{ fanouts.shards.nodes }}           # all fanout members
{{ fanouts.shards.summary.completed }}
{{ item.number }}                    # current fanout member fields
```

## Fanout (Parallel Shards)

`fanout(node, source)` -- source type determines mode:

```python
# int = count (N identical copies)
shards = fanout(codex(task_id="shard", prompt="Shard {{ item.number }}/{{ item.count }}"), 128)

# list = values (one per item)
reviews = fanout(
    codex(task_id="review", prompt="Review {{ item.repo }}"),
    [{"repo": "api"}, {"repo": "billing"}],
)

# dict = matrix (cartesian product)
fuzz = fanout(
    codex(task_id="fuzz", prompt="{{ item.target }} + {{ item.sanitizer }}"),
    {"lib": [{"target": "libpng"}], "check": [{"sanitizer": "asan"}, {"sanitizer": "ubsan"}]},
)
```

### item fields

| Field | Type | Example |
|---|---|---|
| `item.index` | int | 0, 1, 2 |
| `item.number` | int | 1, 2, 3 (1-indexed) |
| `item.count` | int | total copies |
| `item.suffix` | str | "000", "001" (zero-padded) |
| `item.node_id` | str | "shard_001" |
| `item.<key>` | Any | dict keys lifted from values |

### derive (computed fields)

```python
fanout(node, 128, derive={"workspace": "agents/{{ item.suffix }}"})
```

## Merge (Reduce Fanout)

`merge(node, source, by=[...] | size=N)`:

```python
# Batch reduce: one reducer per 16 shards
batch = merge(
    codex(task_id="batch", prompt="Reduce shards {{ item.start_number }}-{{ item.end_number }}"),
    shards, size=16,
)

# Group by field value
family = merge(
    codex(task_id="family", prompt="Reduce {{ item.target }}"),
    fuzz, by=["target"],
)
```

Merge adds: `item.member_ids`, `item.members`, `item.size`, `item.source_group`.
At runtime: `item.scope.nodes`, `item.scope.outputs`, `item.scope.summary`, `item.scope.with_output`.

## Cycles (Iterative Loops)

Use `on_failure` back-edges for retry-until-success patterns:

```python
with Graph("iterative", max_iterations=5) as g:
    write = codex(task_id="write", prompt=(
        "Write the code.\n"
        "{% if nodes.review.output %}Fix: {{ nodes.review.output }}{% endif %}"
    ), tools="read_write")
    review = claude(task_id="review", prompt=(
        "Review:\n{{ nodes.write.output }}\n"
        "If complete, say LGTM. Otherwise list issues."
    ), success_criteria=[{"kind": "output_contains", "value": "LGTM"}])
    done = codex(task_id="done", prompt="Summarize:\n{{ nodes.write.output }}")

    write >> review
    review.on_failure >> write   # loop back until LGTM
    review >> done               # exit on success
```

## Execution Targets

### Local (default)
No `target` needed. Runs on the host machine.

### SSH
```python
target={"kind": "ssh", "host": "server", "username": "deploy"}
# forward_credentials=True to override remote with local codex/claude/kimi auth
target={"kind": "ssh", "host": "server", "forward_credentials": True}
```

### EC2 (auto-discovers AMI, key pair, VPC)
```python
target={"kind": "ec2", "region": "us-east-1"}
# Optional: instance_type, terminate, snapshot, shared, spot
```

### ECS Fargate (auto-discovers VPC, builds agent image)
```python
target={"kind": "ecs", "region": "us-east-1"}
# Optional: image, cpu, memory, install_agents, cluster
```

### Shared instances
Same `shared` ID = same instance across nodes:

```python
plan = codex(task_id="plan", ..., target={"kind": "ec2", "shared": "dev"})
impl = codex(task_id="impl", ..., target={"kind": "ec2", "shared": "dev"})
# Both run on same EC2, files persist between them
```

## Worktrees

Isolate each agent in its own git worktree so they can edit files without conflicts:

```python
with Graph("review", use_worktree=True) as g:
    reviewers = fanout(
        codex(task_id="reviewer", prompt="Review {{ item.file }}", tools="read_write"),
        [{"file": "api.py"}, {"file": "auth.py"}, {"file": "db.py"}],
    )
```

Each agent gets a full repo copy at `.agentflow/worktrees/<run_id>/<node_id>/`. Cleaned up after execution.

## Utility Nodes

Non-LLM nodes for deterministic operations (no API calls, instant execution):

```python
# Run a shell script
build = shell(task_id="build", script="npm run build && echo OK")

# Run Python code
validate = python_node(task_id="validate", code="import json; print(json.dumps({'ok': True}))")

# Sync local repo to remote (rclone or tar+ssh fallback)
deploy = sync(task_id="deploy", mode="full", target={
    "kind": "ssh", "host": "server", "username": "deploy", "remote_workdir": "/app",
})
# mode="repo": .git + stash only (lightweight)
# mode="full": entire directory
```

Mix with agent nodes freely: `build >> codex(...) >> deploy`

## Scratchboard

Enable shared memory across all agents:

```python
with Graph("campaign", scratchboard=True) as g:
    ...
```

All agents get a `scratchboard.md` file to read context and write findings.

## Graph Options

```python
Graph("name",
    concurrency=4,          # max parallel nodes
    fail_fast=False,         # skip downstream on failure
    max_iterations=10,       # cycle iteration limit
    scratchboard=False,      # shared memory file
    use_worktree=False,      # git worktree per agent
    node_defaults={...},     # defaults for all nodes
    agent_defaults={...},    # per-agent defaults
)
```

## CLI

```bash
agentflow run pipeline.py                # run pipeline
agentflow run pipeline.py --output summary
agentflow inspect pipeline.py            # show graph structure
agentflow validate pipeline.py           # check without running
agentflow templates                       # list starter templates
agentflow init > pipeline.py             # scaffold starter
```
登录后下载/点赞/收藏 ❤ 170 | ★ 0
agentflow

目录结构

SKILL.md

举报内容

提示
