Coding Agent (AWS Bedrock)

A readline-based coding agent that runs Claude through Bedrock with configurable authentication and Anthropic features.

examples/coding-agent-bedrock.rs documents the Bedrock-specific provider variant in Appam. The sample supports both SigV4 and bearer-token authentication, then applies Anthropic controls such as adaptive thinking, effort, and retries on top of the Bedrock transport.

Run with SigV4

export AWS_ACCESS_KEY_ID="your-access-key"
export AWS_SECRET_ACCESS_KEY="your-secret-key"
export AWS_SESSION_TOKEN="your-session-token" # optional
export AWS_REGION="us-east-1"                 # optional
export AWS_BEDROCK_MODEL_ID="us.anthropic.claude-opus-4-6-v1" # optional

cargo run --example coding-agent-bedrock

Run with bearer-token auth

export AWS_BEARER_TOKEN_BEDROCK="your-token"
export AWS_BEDROCK_AUTH_METHOD="bearer_token"
export AWS_REGION="us-east-1"

cargo run --example coding-agent-bedrock

What this example actually configures

LlmProvider::Bedrock { region, model_id, auth_method }
ThinkingConfig::adaptive() for the default Opus 4.6 Bedrock model
Anthropic EffortLevel::Max
Prompt caching is supported by Appam for compatible Bedrock Claude models, but this example leaves it off because AWS support is model-specific and the default Opus 4.6 example model is not documented on the current Bedrock prompt-caching support table.
Anthropic beta features for context_1m and effort
Anthropic retry configuration
The shared coding tools: read_file, write_file, bash, and list_files

Key builder setup

let agent = AgentBuilder::new("claude-bedrock")
    .provider(LlmProvider::Bedrock {
        region: region.clone(),
        model_id: model_id.clone(),
        auth_method: bedrock_auth,
    })
    .model(&model_id)
    .system_prompt(
        "You are an advanced coding assistant powered by Claude Opus 4.6 via AWS Bedrock. \
         You have access to file operations, bash commands, and directory listing. \
         Use your adaptive thinking capabilities to reason through complex problems. \
         Always explain your reasoning process and provide detailed analysis.",
    )
    .thinking(appam::llm::anthropic::ThinkingConfig::adaptive())
    .effort(appam::llm::anthropic::EffortLevel::Max)
    .beta_features(appam::llm::anthropic::BetaFeatures {
        context_1m: true,
        effort: true,
        ..Default::default()
    })
    .tool_choice(appam::llm::anthropic::ToolChoiceConfig::Auto {
        disable_parallel_tool_use: false,
    })
    .retry(appam::llm::anthropic::RetryConfig {
        max_retries: 5,
        initial_backoff_ms: 2000,
        max_backoff_ms: 60000,
        backoff_multiplier: 2.0,
        jitter: true,
    })
    .with_tool(Arc::new(read_file()))
    .with_tool(Arc::new(write_file()))
    .with_tool(Arc::new(bash()))
    .with_tool(Arc::new(list_files()))
    .max_tokens(20000)
    .build()?;

Runtime behavior

SigV4 is the default and the example labels it as the streaming-capable path.
Setting AWS_BEDROCK_AUTH_METHOD=bearer_token switches to BedrockAuthMethod::BearerToken.
If you switch to a Bedrock Claude model that supports prompt caching, you can enable .caching(...) and Appam will emit Bedrock-compatible Anthropic cache checkpoints.
The terminal UI prints reasoning, tool calls, and tool results as the stream progresses.