LiteAgents 0.3.0a6: portable tools and simpler setup

LiteAgents keeps one application setup while you change the harness. This preview adds simpler tools and scoped native controls, and updates every cookbook to use the same direct-provider setup.

• Pass typed Python functions as tools; LiteAgents generates their schemas and adapts them to each harness.
• Use LiteAgentClient(profile=profile, tools=[lookup_order]) and query(prompt=..., profile=profile) directly. Existing LiteAgentOptions and Tool classes remain supported.
• Declare subagents without a separate enable flag.
• Put optional native controls under each harness's name in harness_options; switching and fallback select the matching settings automatically.
• Follow short Colab walkthroughs for tools, MCP, conversations, approvals, retries, fallback, profiles and Temporal recovery. Provider keys work directly through the LiteLLM SDK; a gateway is optional.

Install the preview with the harnesses you want:

python -m pip install "liteagents[pydantic-ai,claude-sdk] @ https://github.com/BerriAI/liteagents/releases/download/v0.3.0a6/liteagents-0.3.0a6-py3-none-any.whl"

Getting started · Cookbooks · SDK contract

Validation exercises the same function-and-MCP application across all six harness selectors in direct and Temporal execution, with real native loops and controlled model responses. Fresh-wheel checks include notebook reruns and a separate Temporal worker crash/restart. These are not live-provider acceptance tests.

For durable deployments, upgrade clients and workers together and use a new profile version for new jobs. Finish existing jobs on their original SDK/profile versions. Interrupted external writes still require application idempotency.

LiteAgents 0.3.0a5: guided feature Colabs and recovery fix

The feature Colabs now teach one task at a time: install, enter a key, and run the actual SDK calls. Each explains the expected result and what to change next. The cookbook index follows a learning progression from your first agent to tools, MCP, harness switching, and optional Temporal recovery.

This preview also fixes Pydantic AI checkpoint replay after worker restart. Fresh native conversation IDs no longer cause completed model requests and tools to run again. IDs inside application inputs and outputs are retained.

Durable deployments: upgrade clients and workers together, use a new profile version for new jobs, and finish existing jobs on their original SDK version. Stored request fingerprints change in this preview.

Try the Python tools Colab · All walkthrough previews and validation

Install the preview and the two introductory harnesses:

python -m pip install "liteagents[pydantic-ai,claude-sdk] @ https://github.com/BerriAI/liteagents/releases/download/v0.3.0a5/liteagents-0.3.0a5-py3-none-any.whl"

The PyPI project currently named liteagents is a different package. These examples use the LiteLLM SDK with your model-provider key; neither a gateway nor Temporal is required for ordinary agents.

Validation uses real harnesses, MCP processes, tools, workers, and controlled model responses. It includes repeated notebook runs, fresh wheel installation, and worker crash/restart recovery. It does not establish acceptance for every live model/provider.

LiteAgents 0.3.0a4: a simpler first run

Use await run(prompt, profile=profile) to run an independent task and read its answer from result.text. Change profile.harness to use another harness with the same model and application code. Existing query(), conversation clients, and run controls remain available.

The README, getting-started guide, and first Colab now introduce this simpler flow. The notebook has four steps: install, enter a key, run, and switch harnesses.

Try the updated Colab · Change and validation details

Install the preview and both example harnesses:

python -m pip install "liteagents[pydantic-ai,claude-sdk] @ https://github.com/BerriAI/liteagents/releases/download/v0.3.0a4/liteagents-0.3.0a4-py3-none-any.whl"

The PyPI project currently named liteagents is a different package. This GitHub preview wheel contains the new API. Model-provider credentials are required for your runs; neither a gateway nor Temporal is required.

Validated with all six installed harness selectors and controlled model responses, real tool loops, Temporal workers, repeated notebook execution, and a clean wheel installation. These checks do not establish live acceptance for every provider/model.

LiteAgents 0.3.0a3: standalone Colab cookbooks

LiteAgents 0.3.0a3 adds a simpler SDK quickstart and 12 standalone Google Colab cookbooks. Start with a provider key, a profile, and query(); no checkout or gateway is required.

• Copyable examples for OpenAI, Anthropic, OpenRouter, and more providers, with separate cloud/local authentication guidance.
• Colab examples for conversations, tools, MCP, harness switching, coding comparisons, approvals, subagents, retries, model fallback, and optional Temporal recovery.
• Fix stdio MCP startup in notebook kernels whose stderr has no OS file descriptor.

Install with Python 3.11+:

python -m pip install \
  'liteagents[deepagents] @ https://github.com/BerriAI/liteagents/releases/download/v0.3.0a3/liteagents-0.3.0a3-py3-none-any.whl' \
  -c https://github.com/BerriAI/liteagents/releases/download/v0.3.0a3/constraints-tested.txt

Install only the harness integrations you use, or replace [deepagents] with [all]. OpenCode also needs npm install -g opencode-ai@1.18.29; the Colab cells handle that when selected. The PyPI project currently named liteagents is a different package, so use this preview wheel.

Start in Colab · Getting started · All cookbooks · Model setup

The gateway remains optional: pass its URL, exact model alias, and key in your profile. Special environment variables are not required. Temporal is also optional; its Colab demo uses temporary runtime storage. Persistent applications need an external Temporal service and durable checkpoint storage.

Validation includes real notebook kernels, repeated runs outside the checkout, all six installed native harness selectors, MCP subprocesses, approvals, tool retries, and Temporal worker crashes. A clean install of the built wheel ran the first-agent example on all six harnesses. Provider responses in these checks are deterministic fixtures, not new live-provider acceptance for every documented model. See PR #13 and the validation guide.

All seven PR CI jobs passed, including Python 3.11–3.13 with MCP 1.x/2.x and the native/Temporal/PostgreSQL deployment job. The full integration suite finished with 445 passed and 25 opt-in/environment-dependent skips.

The public client/profile API is unchanged. Upgrade durable clients and workers together.

LiteAgents 0.3.0a2: exact gateway model aliases

LiteAgents 0.3.0a2 lets you use your gateway's model alias directly. Set the endpoint, key, and exact model name, then change the harness without changing the model connection.

from liteagents import ProfileOptions

profile = ProfileOptions(
    harness="deepagents",
    model="my-model",
    model_kwargs={
        "api_base": "https://your-gateway.example/v1",
        "api_key": "your-gateway-key",
    },
)

Aliases containing slashes are sent unchanged. Existing litellm_proxy/alias configurations continue to work. Without a gateway endpoint, use a LiteLLM provider/model name and provider credentials. Both paths go through LiteLLM.

Custom native provider endpoints: if you previously supplied api_base to override a provider's native API endpoint, add custom_llm_provider to model_kwargs, for example "anthropic" or "openai". Without this override, an endpoint-configured name such as anthropic/foo is now the literal gateway alias. Gateway users need no additional setting. See the migration guide.

Install with Python 3.11+ (3.12 recommended):

python -m pip install \
  'liteagents[deepagents] @ https://github.com/BerriAI/liteagents/releases/download/v0.3.0a2/liteagents-0.3.0a2-py3-none-any.whl' \
  -c https://github.com/BerriAI/liteagents/releases/download/v0.3.0a2/constraints-tested.txt

Install only the harnesses you use, or replace [deepagents] with [all] to compare them. OpenCode also requires npm install -g opencode-ai@1.18.29. A simple agent needs neither Temporal nor PostgreSQL. Use this GitHub package; the PyPI project named liteagents is a different package.

The README, JSON/YAML profiles, and cookbooks now use plain gateway aliases. Getting started · Profiles · Cookbooks

Local regression: 286 passed, 8 environment-dependent skips, and 23 opt-in live-provider cases deselected. Validation includes 60 cross-harness tests using real installed runtimes with a deterministic local model endpoint: direct/Temporal execution, tools, MCP, conversations, and tool-free runs. HTTP tests verify plain and slash-containing aliases, streaming, legacy prefixes, and explicit native-provider endpoints. Wheel/source builds, installed-wheel smoke checks, lint, types, and source-size checks pass. These checks do not claim new live-provider acceptance coverage. All seven CI jobs passed, including the Python/MCP matrix, native runtimes, Temporal, PostgreSQL, and deployment worker image build. See PR #11 for CI results.

Upgrade clients and workers together. Complete existing durable runs with their original SDK/profile version and use new profile IDs for changed configuration.

LiteAgents 0.3.0a1: harness portability preview

LiteAgents 0.3.0a1 makes the shared profile portable across DeepAgents, Pydantic AI, Claude Agent SDK, Codex, and both OpenCode adapters. Change the harness while keeping your model configuration, application tools, MCP servers, and client code. LiteLLM translates model requests inside the SDK, and compatible native controls remain available through harness_options.

Install this preview with Python 3.11+ (3.12 recommended):

python -m pip install \
  'liteagents[deepagents] @ https://github.com/BerriAI/liteagents/releases/download/v0.3.0a1/liteagents-0.3.0a1-py3-none-any.whl' \
  -c https://github.com/BerriAI/liteagents/releases/download/v0.3.0a1/constraints-tested.txt

A simple agent needs neither Temporal nor PostgreSQL. Replace [deepagents] with [all] to install the other Python harnesses and MCP/Temporal support. OpenCode additionally requires npm install -g opencode-ai@1.18.29. PostgreSQL is a separate postgres extra. Use this GitHub package; the PyPI project named liteagents is a different package.

Repeated query() calls on one client share conversation history in direct and Temporal execution. start_run() creates independent jobs in both modes. Durable follow-ups store immutable history snapshots in SDK storage; Temporal receives a reference. Recorded operations and checkpoints recover after worker loss. Interrupted external effects still require application idempotency. Existing native sessions cannot migrate between harnesses.

Upgrading from 0.2.0: omitted profile.tools now exposes registered application and discovered MCP tools only; select workspace tools explicitly. An empty list disables tools. Direct start_run() jobs no longer share conversation history; use query() for follow-ups. Finish existing durable runs with their original SDK and profile version, and upgrade clients and workers together. See the migration guide.

Getting started · JSON/YAML profiles · Harness-switching cookbook · All cookbooks · Validation

Validation covers six Python/MCP dependency combinations, the native harness/Temporal/PostgreSQL suite, the self-hosted deployment worker image, 23 live-provider checks using one unchanged model alias, all six harness selectors with tools and MCP in direct and Temporal execution, worker-crash recovery, and fresh-wheel examples without Temporal or PostgreSQL installed. SHA256SUMS covers the attached wheel, source distribution, and tested dependency constraints.

Implementation: #9

LiteAgents 0.2.0

LiteAgents 0.2.0 replaces the original public SDK with profiles and native agent harnesses: DeepAgents, Pydantic AI, Claude Agent SDK, Codex, and both OpenCode API generations.

Use one client for shared application tools, MCP, streaming, subagents, approvals, retries, and optional durable execution through Temporal. A simple agent requires neither Temporal nor PostgreSQL.

Install with Python 3.11+ (3.12 recommended):

python -m pip install 'liteagents[deepagents] @ https://github.com/BerriAI/liteagents/releases/download/v0.2.0/liteagents-0.2.0-py3-none-any.whl'

Replace [deepagents] with [all] to install every Python harness and Temporal/MCP support. OpenCode additionally requires npm install -g opencode-ai@1.18.29; PostgreSQL support is a separate postgres extra. Use this release URL because the existing PyPI liteagents listing is a different package.

Getting started · JSON/YAML profiles · Cookbooks · Migration · Validation

This release defines explicit tool selection (None for defaults, [] for no tools), adapts shared tools/MCP automatically, normalizes tool names, and reports capabilities for the effective execution mode. Native model protocols and supported settings still differ by harness.

Direct queries share conversation history; Temporal submissions are independent jobs. Recovery reuses completed recorded operations; interrupted external effects require application idempotency. Finish existing durable runs on their original SDK version, and upgrade clients and workers together. The former API remains available under liteagents.legacy with the legacy extra.

Validation includes the full native/Temporal/PostgreSQL regression suite, live-provider checks across all six harnesses, real MCP transports, worker-crash recovery, and clean package installation. SHA256SUMS covers the attached wheel and source distribution.