Skip to content
Chimera

Build Your Own Coding Agent in 50 Lines

Build Your Own Coding Agent in 50 Lines

Section titled “Build Your Own Coding Agent in 50 Lines”

This tutorial walks you through building a working coding agent — one that reads files, edits code, runs tests, and uses bash — using Chimera. By the end you’ll have a CLI tool you can point at any repo.

Prerequisites

Section titled “Prerequisites”
Terminal window
git clone https://github.com/0bserver07/chimera && cd chimera
uv sync --extra dev --extra anthropic


# Set up your LLM provider
export ANTHROPIC_API_KEY="sk-ant-..."  # or any Anthropic-compatible endpoint

Step 1: The Minimal Agent (4 lines)

Section titled “Step 1: The Minimal Agent (4 lines)”
my_agent.py
import chimera


provider = chimera.create_provider()
agent = chimera.Agent(provider=provider, tools=list(chimera.AGENT_TOOLS))
result = agent.run("What files are in this directory?", env=chimera.LocalEnvironment("."))
print(result.output)

Run it:

Terminal window
python my_agent.py

That’s a working agent. It has 15 tools (file read/write/edit, bash, search, git, test, web search, etc.), a ReAct loop, and auto-detected provider. But it’s one-shot — let’s make it interactive.

Step 2: Interactive REPL (15 lines)

Section titled “Step 2: Interactive REPL (15 lines)”
my_repl.py
import chimera


provider = chimera.create_provider()
env = chimera.LocalEnvironment(".")
env.setup()


agent = chimera.Agent(
    provider=provider,
    tools=list(chimera.AGENT_TOOLS),
    loop=chimera.ReAct(max_steps=30),
)


while True:
    task = input("\n> ").strip()
    if task in ("exit", "quit"):
        break
    result = agent.run(task, env=env)
    print(result.output)


env.cleanup()

Now you have a conversational coding agent. Ask it to fix bugs, write tests, refactor code.

Step 3: Add Infrastructure (35 lines)

Section titled “Step 3: Add Infrastructure (35 lines)”

A real coding agent needs cost tracking, event monitoring, and undo:

my_coding_agent.py
import chimera


provider = chimera.create_provider()
env = chimera.LocalEnvironment(".")
env.setup()


# Infrastructure
event_bus = chimera.EventBus()
event_bus.subscribe("tool_call", lambda e: print(f"  [{e.metadata.get('tool_name', '?')}]"))
cost_tracker = chimera.CostTracker()
ghost = chimera.GhostCommitManager(".")


config = chimera.LoopConfig(
    event_bus=event_bus,
    cost_tracker=cost_tracker,
    ghost_commits=ghost,
    truncation=chimera.TruncationConfig(max_lines=200),
)


agent = chimera.Agent(
    provider=provider,
    tools=list(chimera.AGENT_TOOLS),
    loop=chimera.ReAct(max_steps=30, config=config),
    prompt=chimera.Prompt.from_string(
        "You are an expert coding agent. Read code before editing. "
        "Run tests after changes. Be precise and minimal."
    ),
)


while True:
    task = input("\n> ").strip()
    if task == "exit":
        break
    if task == "undo":
        print(ghost.undo())
        continue
    if task == "cost":
        print(f"${cost_tracker.total_cost:.4f}")
        continue
    result = agent.run(task, env=env)
    print(result.output)


env.cleanup()

Now you have:

  • Live tool call display (see what the agent does in real-time)
  • Cost tracking (know how much each task costs)
  • Undo (revert the last edit with “undo”)
  • Output truncation (long bash output won’t flood the context)

Step 4: Use a Preset (4 lines)

Section titled “Step 4: Use a Preset (4 lines)”

Instead of configuring everything, use a named preset on CodingAgent:

from chimera.assembly.coding_agent import CodingAgent


# SWE-bench style: minimal scaffold, root-cause focus
agent = CodingAgent.from_preset("swebench")


# Codex style: full tool suite
agent = CodingAgent.from_preset("codex")


# Kimi style: action-first, KISS, iterate on failures
agent = CodingAgent.from_preset("kimi")


# Default: full-featured coding agent
agent = CodingAgent.from_preset("coding_agent")

Migrating from v0.4 / v0.5? Legacy chimera.AgentPreset.* siblings emit a DeprecationWarning and will be removed in v0.7.0. Use CodingAgent.from_preset(...) above. See v0.4 → v0.5 migration.

Step 5: Compose Your Own Architecture

Section titled “Step 5: Compose Your Own Architecture”

Mix and match primitives to create something new:

import chimera
from chimera.core.loops.retry import RetryLoop
from chimera.core.loops.lint_feedback import LintFeedbackLoop


provider = chimera.create_provider()


# Your custom loop: retry with lint checking
inner = chimera.ReAct(max_steps=20)
lint_loop = LintFeedbackLoop(inner=inner, linter="ruff", max_rounds=2)
retry_loop = RetryLoop(inner=lint_loop, max_retries=3)


agent = chimera.Agent(
    provider=provider,
    tools=list(chimera.AGENT_TOOLS),
    loop=retry_loop,
)


result = agent.run("Fix the bug in auth.py", env=chimera.LocalEnvironment("."))

That’s a retry loop wrapping a lint feedback loop wrapping a ReAct loop. If the agent’s fix has lint errors, it fixes them. If the fix still doesn’t work, it retries from scratch. Three levels of resilience, composed from modular blocks.

What’s Next

Section titled “What’s Next”