Agent Loop

Overview

The agent loop is the heart of klaw. It orchestrates the conversation between user input, LLM decisions, and tool execution, with built-in safety guardrails for iteration limits, context management, cost tracking, and human approval.

The Loop

┌──────────────────────────────────────────────────────────────────────┐
│                           AGENT LOOP                                │
│                                                                     │
│  ┌──────────┐    ┌──────────┐    ┌──────────┐    ┌──────────────┐  │
│  │ Receive  │ →  │ Planning │ →  │  Context │ →  │     LLM      │  │
│  │  Input   │    │  (opt.)  │    │  Check   │    │   Decision   │  │
│  └──────────┘    └──────────┘    └──────────┘    └──────┬───────┘  │
│       ▲                                                  │         │
│       │                                                  ▼         │
│  ┌──────────┐    ┌──────────┐    ┌──────────┐    ┌──────────────┐  │
│  │  Cost &  │ ←  │ Reflect  │ ←  │ Execute  │ ←  │   Approval   │  │
│  │ Iter Chk │    │  (opt.)  │    │  Tools   │    │   (opt.)     │  │
│  └──────────┘    └──────────┘    └──────────┘    └──────────────┘  │
│       │                                                            │
│       └── Loop or Complete ─────────────────────────────────────── │
└──────────────────────────────────────────────────────────────────────┘

Step-by-Step Flow

Receive Input

User sends a message through a channel (CLI, Slack, API).

msg, _ := channel.Receive(ctx)
history = append(history, Message{Role: "user", Content: msg})

Planning Phase (Optional)

On the first message, if planning is enabled, a planning prompt is injected asking the LLM to outline steps before acting.

if planner.Enabled && isFirstMessage {
    history = InjectPlanRequest(history, planner.PlanPrompt)
}

The default planning prompt asks the LLM to:

Analyze what the user is asking for
List steps (max 5)
Identify potential issues

Context Window Check

Before each LLM call, the context manager estimates token usage. If the conversation exceeds the compaction threshold, middle messages are summarized via an LLM call.

history, _ = contextMgr.CompactIfNeeded(ctx, provider, history)

Max context: 200,000 tokens (default)
Compaction threshold: 75% of available context
Reserve: 8,192 tokens for the response
Keeps first user message and last 6 messages verbatim

Send to LLM

Agent sends conversation history to the LLM provider with resilient delivery (retry + fallback).

events, _ := provider.Stream(ctx, history)

Process Stream

Agent collects streaming events: text chunks, tool calls, stop signals with usage data.

for event := range events {
    switch event.Type {
    case "text":
        channel.Send(ctx, Message{Content: event.Text, Partial: true})
    case "tool_use":
        toolCalls = append(toolCalls, event.ToolCall)
    case "stop":
        costTracker.Add(model, event.Usage.InputTokens, event.Usage.OutputTokens)
        break
    }
}

Human Approval (Optional)

If a tool is in the require_approval list, the user is prompted before execution.

if approval.NeedsApproval(call.Name) {
    approved := approval.RequestApproval(ctx, channel, call)
    if !approved {
        result = "Denied by user"
        continue
    }
}

The user sees: ⚠ Tool 'bash' requires approval. Execute? [y/N]:

Execute Tools (Parallel)

Approved tools execute concurrently. Each tool runs in its own goroutine with a 2-minute timeout. Results are collected in original call order after all goroutines complete.

// All approved tools run in parallel
var wg sync.WaitGroup
results := make([]*tool.Result, len(toolCalls))
for i, call := range approved {
    wg.Add(1)
    go func(idx int, tc ToolCall) {
        defer wg.Done()
        ctx, cancel := context.WithTimeout(ctx, 2*time.Minute)
        results[idx] = tools.Execute(ctx, tc.Name, tc.Input)
        cancel()
    }(i, call)
}
wg.Wait()

If the LLM returns 3 file reads simultaneously, they complete in roughly the time of one read instead of three sequential reads.

Reflection (Optional)

After every N tool calls (default: 3), a reflection prompt is injected asking the LLM to assess progress.

if ShouldReflect(reflectionConfig, toolCallsSinceReflection) {
    history = InjectReflection(history, reflectionConfig.Prompt)
    toolCallsSinceReflection = 0
}

Cost & Iteration Check

The loop checks two safety limits before continuing:

// Iteration limit (default: 50)
if iteration >= maxIterations {
    return AgentError{Code: ErrMaxIterations}
}

// Cost budget
if err := costTracker.CheckBudget(); err != nil {
    return AgentError{Code: ErrBudgetExceed}
}

Loop or Complete

If tools were called, loop back to the context check step. If no tools were requested, the response is complete.

if len(toolCalls) > 0 {
    continue // Back to Step 3
}
channel.Send(ctx, Message{Content: text, Done: true})

Iteration Limit

Each agent has a maximum number of loop iterations (default: 50). This prevents runaway loops where the LLM repeatedly calls tools without converging on a solution.

[agent.default]
max_iterations = 50

[agent.researcher]
max_iterations = 30

When the limit is reached, the agent stops with AgentError{Code: ErrMaxIterations}.

Context Window Management

Large conversations can exceed the LLM’s context window. The context manager handles this automatically:

Parameter	Default	Description
`MaxContextTokens`	200,000	Maximum tokens before compaction
`CompactionThreshold`	0.75	Ratio of max tokens that triggers compaction
`ReserveTokens`	8,192	Tokens reserved for LLM response

Compaction strategy:

Keep the first user message verbatim (preserves original intent)
Keep the last 6 messages verbatim (preserves recent context)
Summarize everything in between via an LLM call
Replace middle messages with [Previous conversation summary]

Token estimation uses a ~4 characters per token heuristic.

Cost Tracking

Every LLM call records input and output token counts. The cost tracker computes session cost using per-model pricing tables:

Model	Input (per 1M)	Output (per 1M)
`claude-sonnet-4-20250514`	$3.00	$15.00
`claude-opus-4-20250514`	$15.00	$75.00
`claude-3-5-haiku-20241022`	$0.80	$4.00
`openai/gpt-4o`	$2.50	$10.00
`deepseek/deepseek-chat`	$0.14	$0.28

Set a budget with max_session_cost in config. When the budget is reached, the agent stops with ErrBudgetExceed. A warning is logged at 80% of the budget.

Structured Errors

The agent loop uses typed errors with machine-readable codes:

Error Code	Constant	Description
`max_iterations`	`ErrMaxIterations`	Reached iteration limit
`provider_error`	`ErrProvider`	LLM API or provider failure
`tool_execution`	`ErrToolExec`	Tool execution failed
`context_limit`	`ErrContextLimit`	Context window exceeded
`budget_exceeded`	`ErrBudgetExceed`	Session cost budget exceeded

type AgentError struct {
    Code    ErrorCode  // Machine-readable code
    Message string     // Human-readable message
    Cause   error      // Underlying error (optional)
}

Errors format as [CODE] message: cause and support errors.Unwrap().

Stream Events

The provider returns a stream of events:

Event Type	Description	Fields
`text`	Text chunk to stream to user	`Text`
`tool_use`	LLM wants to call a tool	`ToolCall`
`tool_result`	Result of a tool call	`Content`
`error`	Error occurred	`Error`
`stop`	Generation complete	`Usage{InputTokens, OutputTokens}`

The stop event now includes a Usage field with token counts, used by the cost tracker.

Tool Execution

Timeout

Each tool has a 2-minute timeout by default:

ctx, cancel := context.WithTimeout(ctx, 2*time.Minute)
defer cancel()
result, err := tool.Execute(ctx, input)

Result Formatting

Tool results are formatted for display:

╭──────────────────────────────────────────╮
│ Tool: bash                               │
│ Command: git status                      │
├──────────────────────────────────────────┤
│ On branch main                           │
│ Your branch is up to date                │
│ nothing to commit, working tree clean    │
╰──────────────────────────────────────────╯

Error Handling

Tool errors are captured and fed back to the LLM as structured errors:

result, err := tool.Execute(ctx, input)
if err != nil {
    return AgentError{Code: ErrToolExec, Message: err.Error(), Cause: err}
}

Memory Integration

Before each LLM call, workspace context is loaded:

System Prompt = Base Instructions
             + SOUL.md content
             + AGENTS.md content
             + TOOLS.md content
             + Skill prompts

Sub-Agent Delegation

The delegate tool allows the main agent to spawn ephemeral sub-agents that run inline during a conversation. This is useful for parallelizable sub-tasks or specialized work.

┌───────────────────────────────────────────────┐
│              Main Agent                        │
│  "Research patterns and summarize"             │
│                                                │
│  ┌─────────────┐    ┌─────────────┐           │
│  │  delegate    │    │  delegate    │           │
│  │  (depth 1)  │    │  (depth 1)  │           │
│  │  research   │    │  summarize  │           │
│  └──────┬──────┘    └──────┬──────┘           │
│         │                   │                  │
│         │  can delegate     │                  │
│         ▼                   ▼                  │
│  ┌─────────────┐    ┌─────────────┐           │
│  │  delegate    │    │  (result)   │           │
│  │  (depth 2)  │    └─────────────┘           │
│  └─────────────┘                               │
└───────────────────────────────────────────────┘

Key properties:

Sub-agents use RunOnce — a non-streaming agent loop with parallel tool execution
Each sub-agent gets its own tool registry (parent delegate excluded, child delegate added at depth+1)
Maximum nesting depth: 3 levels
5-minute timeout per delegation
Output truncated at 30,000 characters

// The delegate tool calls RunOnce with a callback to avoid circular imports
delegateTool := tool.NewDelegateTool(
    func(ctx context.Context, cfg tool.RunConfig) (string, error) {
        return agent.RunOnce(ctx, agent.RunOnceConfig{
            Provider: cfg.Provider.(provider.Provider),
            Tools:    cfg.Tools,
            Prompt:   cfg.Prompt,
        })
    },
    provider, tools,
)

Performance Considerations

Factor	Impact	Mitigation
Tool execution time	Blocks loop	Timeout, parallel execution
Large outputs	Token cost	Output truncation
Many tool calls	Latency	Parallel execution
Long history	Context limits	Automatic compaction
Runaway loops	Cost & time	Iteration limit
Provider outages	Availability	Retry + fallback
Sub-agent depth	Complexity	Depth limit (max 3)

Next Steps

Resilience

Provider retry, fallback, and error handling

Cost & Safety

Budgets, approval, and monitoring

Tools Reference

Available built-in tools

Custom Tools

Build your own tools

Getting Started

Core Concepts

Guides

Configuration

Architecture

CLI Reference

Resources

Overview

The Loop

Step-by-Step Flow

Iteration Limit

Context Window Management

Cost Tracking

Structured Errors

Stream Events

Tool Execution

Timeout

Result Formatting

Error Handling

Memory Integration

Sub-Agent Delegation

Performance Considerations

Next Steps

Resilience

Cost & Safety

Tools Reference

Custom Tools

Getting Started

Core Concepts

Guides

Configuration

Architecture

CLI Reference

Resources

​Overview

​The Loop

​Step-by-Step Flow

​Iteration Limit

​Context Window Management

​Cost Tracking

​Structured Errors

​Stream Events

​Tool Execution

​Timeout

​Result Formatting

​Error Handling

​Memory Integration

​Sub-Agent Delegation

​Performance Considerations

​Next Steps

Resilience

Cost & Safety

Tools Reference

Custom Tools

Overview

The Loop

Step-by-Step Flow

Iteration Limit

Context Window Management

Cost Tracking

Structured Errors

Stream Events

Tool Execution

Timeout

Result Formatting

Error Handling

Memory Integration

Sub-Agent Delegation

Performance Considerations

Next Steps