Vibe-Coding & AI-Assisted Software Engineering

"The best code is the code you never had to write yourself. The second best is code you wrote with an AI that understood your codebase."

Deploy, Pair-Programming AI Agent

1. Code Completion and Fill-in-the-Middle

The simplest form of AI-assisted coding is inline completion: the developer writes code, and the model predicts what comes next. Modern code completion goes beyond simple next-token prediction with Fill-in-the-Middle (FIM), where the model sees both the code before and after the cursor position. This allows it to generate code that fits seamlessly into existing context rather than just appending to the end.

FIM Architecture

FIM works by rearranging the input during training. A code file is split into a prefix (before cursor), a middle (the target), and a suffix (after cursor). The model receives <PRE> prefix <SUF> suffix <MID> and learns to predict the middle section. This format, sometimes called PSM (Prefix-Suffix-Middle), teaches the model to generate code that is syntactically and semantically consistent with both surrounding contexts. Models like StarCoder, DeepSeek-Coder, and Codestral are trained with FIM from the start. Figure 28.1.1 illustrates the Fill-in-the-Middle approach. Code Fragment 28.1.2 below puts this into practice.

Figure 28.1.2: Fill-in-the-Middle (FIM). The model receives prefix and suffix context, then generates code that fits between them.

# Using a FIM model directly (DeepSeek-Coder example)
from transformers import AutoTokenizer, AutoModelForCausalLM

model_id = "deepseek-ai/deepseek-coder-6.7b-base"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id, torch_dtype="auto", device_map="auto")

# FIM format: prefix + suffix, model fills the middle
prefix = """def binary_search(arr, target):
 left, right = 0, len(arr) - 1
 while left <= right:
"""

suffix = """
 if arr[mid] == target:
 return mid
 elif arr[mid] < target:
 left = mid + 1
 else:
 right = mid - 1
 return -1"""

# DeepSeek FIM tokens
fim_input = "<|fim_begin|>" + prefix + "<|fim_hole|>" + suffix + "<|fim_end|>"

inputs = tokenizer(fim_input, return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=50)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

2. AI-Native IDEs and Coding Assistants

The coding assistant landscape has evolved from simple autocomplete plugins into full AI-native development environments. These tools integrate LLM capabilities deeply into the editing experience, providing not just completions but also chat-based code editing, codebase-aware context, and multi-file refactoring.

3. Agentic Coding

Agentic coding represents the next evolution: instead of suggesting completions that a developer accepts or rejects, the AI operates as an autonomous agent that can read files, write code, run tests, debug errors, and iterate until a task is complete. The developer provides a high-level description and the agent handles the implementation details.

How Agentic Coding Tools Work

Agentic coding tools follow a plan-execute-observe loop. The LLM receives a task description and access to tools (file read/write, terminal execution, search). It plans an approach, executes code changes, runs tests to verify, observes errors, and iterates. This is fundamentally the ReAct pattern (Chapter 22) applied to software engineering. The key differentiator between tools is how they manage context, what tools they expose, and how autonomously they operate. Figure 28.1.3 shows the agentic coding loop. Code Fragment 28.1.2 below puts this into practice.

# Simplified agentic coding loop (conceptual)
from openai import OpenAI
import subprocess, json

client = OpenAI()

def coding_agent(task: str, max_iterations: int = 5):
 tools = [
 {"type": "function", "function": {
 "name": "read_file",
 "parameters": {"type": "object", "properties": {"path": {"type": "string"}}}}},
 {"type": "function", "function": {
 "name": "write_file",
 "parameters": {"type": "object", "properties": {
 "path": {"type": "string"}, "content": {"type": "string"}}}}},
 {"type": "function", "function": {
 "name": "run_command",
 "parameters": {"type": "object", "properties": {"cmd": {"type": "string"}}}}},
 ]

 messages = [{"role": "system", "content": "You are a coding agent. Use tools to complete the task."},
 {"role": "user", "content": task}]

 for i in range(max_iterations):
 response = client.chat.completions.create(
 model="gpt-4o", messages=messages, tools=tools
 )
 msg = response.choices[0].message
 messages.append(msg)

 if not msg.tool_calls:
 return msg.content # Task complete

 for tc in msg.tool_calls:
 result = execute_tool(tc.function.name, json.loads(tc.function.arguments))
 messages.append({"role": "tool", "tool_call_id": tc.id, "content": result})

4. Code Generation from Specifications

A new category of tools generates complete applications from high-level descriptions. Bolt.new, Vercel's v0, and Lovable let users describe an app in natural language and receive a working, deployable application. These tools combine LLM code generation with scaffolding templates, component libraries, and deployment infrastructure to bridge the gap between description and running software.

SWE-bench: Evaluating Coding Agents

SWE-bench provides a rigorous benchmark for evaluating coding agents on real software engineering tasks. It collects actual GitHub issues from popular Python repositories along with their corresponding pull requests. The agent receives the issue description and must produce a patch that passes the repository's test suite. SWE-bench Verified (a human-validated subset of 500 problems) is the standard evaluation set, with top agents scoring around 50 to 60% as of early 2026. Code Fragment 28.1.3 below puts this into practice.

# Evaluating an agent on SWE-bench (conceptual)
from swebench.harness.run_evaluation import run_evaluation

results = run_evaluation(
 predictions_path="predictions.json", # Agent's generated patches
 swe_bench_tasks="princeton-nlp/SWE-bench_Verified",
 log_dir="./eval_logs",
 timeout=900,
)

print(f"Resolved: {results['resolved']} / {results['total']}")
print(f"Pass rate: {results['resolved'] / results['total'] * 100:.1f}%")

Exercises