shipitagentDocs

Build an agent

Reasoning

Turn on a model's thinking, control how much it thinks, and show it in your interface.

1 min read

Some models think before they answer, and some of those show their thinking. When a response includes it, shipit emits a reasoning_completed event with the text, so an interface can show a "thinking" panel with no extra work.

Turning it on

Anthropic models think when you ask them to. Current models use adaptive thinking, where the model decides how much to think; effort sets the depth.

python
from shipit_agent.llms import AnthropicChatLLM

llm = AnthropicChatLLM(model="claude-sonnet-5-5", thinking="adaptive", effort="high")

Older Anthropic models take a fixed budget instead: thinking_budget_tokens=4000. Give a budget to a model that rejects budgets and shipit sends adaptive thinking instead. interleaved_thinking=True lets the model think between tool calls, not only at the start of a turn.

With the SDK API: AgentOptions(model="claude-sonnet-5-5", thinking="adaptive", effort="high"), or max_thinking_tokens= for a budget.

OpenAI reasoning models (o-series and gpt-5) reason by default; reasoning_effort sets how much:

python
from shipit_agent.llms import OpenAIChatLLM

llm = OpenAIChatLLM(model="gpt-5.5", reasoning_effort="high")

Through LiteLLM and Bedrock, models that return reasoning_content, such as gpt-oss and DeepSeek R1, show it with no setting.

Showing it

python
for event in agent.stream("Why does the cache miss on every request?"):
    if event.type == "reasoning_completed":
        print("thinking:", event.payload["content"])
    elif event.type == "tool_called":
        print("calling:", event.payload["tool"])
    elif event.type == "text_delta":
        print(event.payload["chunk"], end="", flush=True)

Thinking arrives whole, as one reasoning_completed event per step, before that step's tool calls or answer. The answer itself streams as text_delta.

With the SDK API, thinking is a ThinkingBlock in the AssistantMessage's content.

Things to know

  • Not every model shows its thinking. A model that does not expose it simply emits no reasoning_* events.
  • Thinking costs output tokens. Use high effort where the task needs it, not by default.
  • Simple steps may have no thinking even with it on; that is the model's choice.
  • Thinking is not narration. shipit never presents a model's reasoning as something it did; progress_summaries=True adds short progress events built from what actually happened.