Building with Pydantic AI Harness
Pydantic AI Harness is the official capability library for Pydantic AI. Capabilities that need model or
framework support -- and those fundamental to every agent -- live in core pydantic-ai; optional,
batteries-included capabilities live here. Both are composed onto an agent through the same
capabilities=[...] API.
This skill covers the capabilities shipped by pydantic-ai-harness. For the core framework -- agents,
tools, structured output, hooks, and testing -- use the building-pydantic-ai-agents skill instead.
When to Use This Skill
Invoke this skill when:
- The user mentions
pydantic-ai-harness, CodeMode, code mode, or the Monty sandbox
- An agent makes many sequential tool calls that could collapse into one sandboxed Python execution
- The user wants the model to write Python that loops, branches, aggregates, or parallelizes tool calls with
asyncio.gather
- The user asks to sandbox or constrain the code an agent runs
Do not use this skill for:
- Core Pydantic AI usage -- building agents, adding tools, structured output, streaming, or testing (use
building-pydantic-ai-agents)
- Capabilities that ship in core
pydantic-ai, such as web search, tool search, and thinking
- The Pydantic validation library on its own (
pydantic/BaseModel without agents)
Supported Capabilities
CodeMode has a full reference below; it is the flagship capability and the one this skill goes deep on.
The rest ship today and each has its own README with API and examples.
Each capability lives in its own submodule and is imported from there
(from pydantic_ai_harness.<module> import ...). Capabilities are not importable from the top-level
pydantic_ai_harness package by design, so each one keeps its own optional dependencies isolated.
CodeMode, FileSystem, Shell, and ManagedPrompt also have top-level re-exports (importable directly
from pydantic_ai_harness).
APIs are subject to change between releases; breaking changes ship deprecation warnings where practical.
| Capability |
Module |
Description |
CodeMode |
pydantic_ai_harness.code_mode (also top-level) |
Wraps eligible tools into a single sandboxed run_code tool so the model orchestrates them in Python -- see Code Mode |
FileSystem |
pydantic_ai_harness.filesystem (also top-level) |
Read, write, edit, and search files under a root directory, with traversal prevention |
Shell |
pydantic_ai_harness.shell (also top-level) |
Run commands in a subprocess with allowlists, a default denylist, timeouts, and env masking |
ManagedPrompt |
pydantic_ai_harness.logfire (also top-level) |
Back an agent's instructions with a Logfire-managed prompt |
SubAgents |
pydantic_ai_harness.subagents |
Delegate subtasks to specialized child agents |
DynamicWorkflow |
pydantic_ai_harness.dynamic_workflow |
Orchestrate sub-agents from a model-written Python script |
Planning |
pydantic_ai_harness.planning |
Break complex tasks into structured plans before execution |
compaction family (SlidingWindowCompaction, SummarizingCompaction, ...) |
pydantic_ai_harness.compaction |
Trim or summarize conversation history to stay within token limits |
ToolOutputLimits |
pydantic_ai_harness.tool_output_limits |
Truncate, summarize, or spill large tool outputs |
RepoContext |
pydantic_ai_harness.repo_context |
Auto-load CLAUDE.md/AGENTS.md and repo structure |
StepPersistence |
pydantic_ai_harness.step_persistence |
Save, restore, resume, and fork run state |
PydanticAIDocs |
pydantic_ai_harness.pydantic_ai_docs |
On-demand read_pyai_docs tool for Pydantic AI docs |
CapabilityCreation |
pydantic_ai_harness.capability_creation |
Let an agent author, validate, and load real capabilities at runtime |
| media externalization |
pydantic_ai_harness.media |
Offload large BinaryContent to content-addressed stores |
Still experimental: an ACP server adapter, imported from pydantic_ai_harness.experimental.acp. Importing it
emits a HarnessExperimentalWarning.
The full, current list with links and status is in the
capability matrix.
Install
uv add pydantic-ai-harness
Each capability declares its own extra. Code Mode needs the Monty sandbox:
uv add "pydantic-ai-harness[codemode]" # `code-mode` is also accepted as an alias
Requires Python 3.10+ and pydantic-ai-slim>=2.18.0.
Quick Start
A harness capability is added to the agent like any other. Here CodeMode wraps locally registered tools
into a single run_code tool that the model drives with Python.
from pydantic_ai import Agent
from pydantic_ai_harness import CodeMode
agent = Agent('anthropic:claude-sonnet-4-6', capabilities=[CodeMode()])
@agent.tool_plain
def get_temperature_f(city: str) -> float:
return {'Paris': 68.0, 'Tokyo': 77.0}[city]
@agent.tool_plain
def convert_temp(fahrenheit: float) -> float:
return round((fahrenheit - 32) * 5 / 9, 1)
result = agent.run_sync(
'Compare the weather in Paris and Tokyo, and report both temperatures in Celsius.'
)
print(result.output)
#> Paris is 20.0 C and Tokyo is 25.0 C.
The model writes a single Python script that fetches both temperatures with asyncio.gather and then
converts them -- performing four tool calls across two dependent stages in one run_code invocation.
Key Practices
- Confirm a harness capability is actually needed. If core Pydantic AI tools and capabilities are enough, use the
building-pydantic-ai-agents skill instead -- don't reach for the harness by default.
- Read the reference before writing code. Each capability has its own configuration, constraints, and gotchas -- load the linked reference (e.g. Code Mode) first.
- Install the capability's extra. Importing
CodeMode without pydantic-ai-harness[codemode] raises an ImportError; the Monty sandbox is an optional dependency.
Common Gotchas
native=True tools bypass CodeMode. Provider-native MCP servers and web search execute server-side, so run_code never sees them. Use native=False for client-side dispatch that CodeMode can wrap, but do not treat a remote server as trusted or sandboxed; see the Code Mode trust boundary.- The Monty sandbox is a Python subset. It has no third-party imports and only a small stdlib allowlist -- read Code Mode before debugging generated code that fails to run.
CodeMode needs its extra. Install pydantic-ai-harness[codemode], not the bare package.
1---2name: pydantic-ai-harness3description: Extend Pydantic AI agents with batteries-included capabilities from pydantic-ai-harness -- Code Mode (collapse many tool calls into one sandboxed Python execution), a filesystem and shell, sub-agents, planning, context compaction, and more. Use when the user mentions pydantic-ai-harness, CodeMode, Monty, code mode, or tool sandboxing, when they want first-party filesystem/shell/sub-agent/planning/compaction capabilities for a Pydantic AI agent, when they want an agent to run agent-written Python, or when a Pydantic AI agent would benefit from orchestrating multiple tool calls in a single sandboxed script.4license: MIT5---6
7# Building with Pydantic AI Harness
8
9Pydantic AI Harness is the official capability library for Pydantic AI. Capabilities that need model or
10framework support -- and those fundamental to every agent -- live in core `pydantic-ai`; optional,
11batteries-included capabilities live here. Both are composed onto an agent through the same
12`capabilities=[...]` API.
13
14This skill covers the capabilities shipped by `pydantic-ai-harness`. For the core framework -- agents,
15tools, structured output, hooks, and testing -- use the `building-pydantic-ai-agents` skill instead.
16
17## When to Use This Skill
18
19Invoke this skill when:
20- The user mentions `pydantic-ai-harness`, `CodeMode`, code mode, or the Monty sandbox
21- An agent makes many sequential tool calls that could collapse into one sandboxed Python execution
22- The user wants the model to write Python that loops, branches, aggregates, or parallelizes tool calls with `asyncio.gather`
23- The user asks to sandbox or constrain the code an agent runs
24
25Do **not** use this skill for:
26- Core Pydantic AI usage -- building agents, adding tools, structured output, streaming, or testing (use `building-pydantic-ai-agents`)
27- Capabilities that ship in core `pydantic-ai`, such as web search, tool search, and thinking
28- The Pydantic validation library on its own (`pydantic`/`BaseModel` without agents)
29
30## Supported Capabilities
31
32`CodeMode` has a full reference below; it is the flagship capability and the one this skill goes deep on.
33The rest ship today and each has its own README with API and examples.
34
35Each capability lives in its own submodule and is imported from there
36(`from pydantic_ai_harness.<module> import ...`). Capabilities are not importable from the top-level
37`pydantic_ai_harness` package by design, so each one keeps its own optional dependencies isolated.
38`CodeMode`, `FileSystem`, `Shell`, and `ManagedPrompt` also have top-level re-exports (importable directly
39from `pydantic_ai_harness`).
40
41APIs are subject to change between releases; breaking changes ship deprecation warnings where practical.
42
43| Capability | Module | Description |
44|---|---|---|
45| `CodeMode` | `pydantic_ai_harness.code_mode` (also top-level) | Wraps eligible tools into a single sandboxed `run_code` tool so the model orchestrates them in Python -- see [Code Mode](./references/CODE-MODE.md) |
46| `FileSystem` | `pydantic_ai_harness.filesystem` (also top-level) | Read, write, edit, and search files under a root directory, with traversal prevention |
47| `Shell` | `pydantic_ai_harness.shell` (also top-level) | Run commands in a subprocess with allowlists, a default denylist, timeouts, and env masking |
48| `ManagedPrompt` | `pydantic_ai_harness.logfire` (also top-level) | Back an agent's instructions with a Logfire-managed prompt |
49| `SubAgents` | `pydantic_ai_harness.subagents` | Delegate subtasks to specialized child agents |
50| `DynamicWorkflow` | `pydantic_ai_harness.dynamic_workflow` | Orchestrate sub-agents from a model-written Python script |
51| `Planning` | `pydantic_ai_harness.planning` | Break complex tasks into structured plans before execution |
52| compaction family (`SlidingWindowCompaction`, `SummarizingCompaction`, ...) | `pydantic_ai_harness.compaction` | Trim or summarize conversation history to stay within token limits |
53| `ToolOutputLimits` | `pydantic_ai_harness.tool_output_limits` | Truncate, summarize, or spill large tool outputs |
54| `RepoContext` | `pydantic_ai_harness.repo_context` | Auto-load CLAUDE.md/AGENTS.md and repo structure |
55| `StepPersistence` | `pydantic_ai_harness.step_persistence` | Save, restore, resume, and fork run state |
56| `PydanticAIDocs` | `pydantic_ai_harness.pydantic_ai_docs` | On-demand `read_pyai_docs` tool for Pydantic AI docs |
57| `CapabilityCreation` | `pydantic_ai_harness.capability_creation` | Let an agent author, validate, and load real capabilities at runtime |
58| media externalization | `pydantic_ai_harness.media` | Offload large `BinaryContent` to content-addressed stores |
59
60Still experimental: an ACP server adapter, imported from `pydantic_ai_harness.experimental.acp`. Importing it
61emits a `HarnessExperimentalWarning`.
62
63The full, current list with links and status is in the
64[capability matrix](https://github.com/pydantic/pydantic-ai-harness#capability-matrix).
65
66## Install
67
68```bash
69uv add pydantic-ai-harness
70```
71
72Each capability declares its own extra. Code Mode needs the Monty sandbox:
73
74```bash
75uv add "pydantic-ai-harness[codemode]" # `code-mode` is also accepted as an alias
76```
77
78Requires Python 3.10+ and `pydantic-ai-slim>=2.18.0`.
79
80## Quick Start
81
82A harness capability is added to the agent like any other. Here `CodeMode` wraps locally registered tools
83into a single `run_code` tool that the model drives with Python.
84
85```python {test="skip"}
86from pydantic_ai import Agent
87
88from pydantic_ai_harness import CodeMode
89
90agent = Agent('anthropic:claude-sonnet-4-6', capabilities=[CodeMode()])
91
92
93@agent.tool_plain
94def get_temperature_f(city: str) -> float:
95 return {'Paris': 68.0, 'Tokyo': 77.0}[city]
96
97
98@agent.tool_plain
99def convert_temp(fahrenheit: float) -> float:
100 return round((fahrenheit - 32) * 5 / 9, 1)
101
102result = agent.run_sync(
103 'Compare the weather in Paris and Tokyo, and report both temperatures in Celsius.'
104)
105print(result.output)
106#> Paris is 20.0 C and Tokyo is 25.0 C.
107```
108
109The model writes a single Python script that fetches both temperatures with `asyncio.gather` and then
110converts them -- performing four tool calls across two dependent stages in one `run_code` invocation.
111
112## Key Practices
113
114- **Confirm a harness capability is actually needed.** If core Pydantic AI tools and capabilities are enough, use the `building-pydantic-ai-agents` skill instead -- don't reach for the harness by default.
115- **Read the reference before writing code.** Each capability has its own configuration, constraints, and gotchas -- load the linked reference (e.g. [Code Mode](./references/CODE-MODE.md)) first.
116- **Install the capability's extra.** Importing `CodeMode` without `pydantic-ai-harness[codemode]` raises an `ImportError`; the Monty sandbox is an optional dependency.
117
118## Common Gotchas
119
120- **`native=True` tools bypass `CodeMode`.** Provider-native MCP servers and web search execute server-side, so `run_code` never sees them. Use `native=False` for client-side dispatch that `CodeMode` can wrap, but do not treat a remote server as trusted or sandboxed; see the [Code Mode trust boundary](./references/CODE-MODE.md#sandbox-restrictions).
121- **The Monty sandbox is a Python subset.** It has no third-party imports and only a small stdlib allowlist -- read [Code Mode](./references/CODE-MODE.md#sandbox-restrictions) before debugging generated code that fails to run.
122- **`CodeMode` needs its extra.** Install `pydantic-ai-harness[codemode]`, not the bare package.
pydantic