How to Design an AI Agent Handoff with Structured Context
Design AI agent handoffs with state, evidence, authority, and recovery without copying the whole history or trusting a vague summary.
Accepting new clients
You run the business.
I handle the software.
I'm Samuel Fajreldines, a senior software engineer. Growing U.S. businesses hire my team to become their software department, not a vendor they rehire every quarter. I set the technical direction, my team builds and ships, and I stay accountable for everything that goes out.
- From $5,000/month
- Ongoing, not a project
- My team handles delivery
Artificial Intelligence
49 articles · practical guides & tutorials by Samuel Fajreldines
In-depth Artificial Intelligence articles, hands-on guides and tutorials by Samuel Fajreldines — 49 expert posts on Artificial Intelligence for software engineers.
Articles in Artificial Intelligence
How to Trace AI Agent Tool Calls with OpenTelemetry
Build a useful AI agent trace that links runs, models, tools, retries, and errors without logging prompts or sensitive data by default during debugging.
AI agent context compaction: what to preserve before continuing
Learn what an AI agent can summarize, what must live outside its history, and how to verify task state after context compaction.
How to Verify an AI Agent Tool Call Before Retrying
Learn how to verify an AI agent tool call after a timeout, record receipts, and retry side-effecting actions without creating duplicates.
How to Pause and Resume an AI Agent Without Restarting
Learn to pause and resume AI agents with checkpoints, durable state, revalidated approvals, and external-effect reconciliation after interruptions safely.
How to Stream AI Agent Progress to a UI
Learn to stream AI agent progress to a UI with typed events, SSE or WebSockets, reconnects, approvals, failures, and a reliable terminal state.
How to Prevent an AI Agent from Looping Forever
Prevent AI agent loops with 4 guardrails: completion, progress, budget, and a terminal state the runtime can explain and recover from in a production system.
How to Avoid Duplicate Tool Calls During LLM Failover
Design LLM retry, failover, and stop decisions without replaying tool calls, losing the output contract, or hiding cost, latency, and partial effects.
How to test an AI agent's trajectory
Learn how to evaluate an AI agent's trajectory by recording tools, arguments, order, state, and interventions before accepting its final answer.
When should an AI agent ask for human approval?
Learn when an AI agent should pause for human approval, how to classify tools by risk, and how to resume the run without duplicating side effects.
How to set a token budget for an AI agent
Learn how to cap tokens, tool calls, and turns per AI agent run with a TypeScript guard that stops before the next model call and records why the run stopped.
How to cancel an AI agent when a tool hangs
Combine AbortSignal, per-step timeouts, and turn limits to safely cancel AI agents, propagate the signal, and avoid unsafe retries after external effects.
Prompt caching for multi-step AI agents with tool calls
Structure tools, system, and history so multi-step agents get cache hits on each tool step. Measure cached tokens, stop prefix invalidation, and cut input cost.
AI software engineer portfolio: what to prove
Build an AI software engineer portfolio that shows the problem, architecture, tests, failures, operations, and decisions behind a project.
AI agent memory: Graphiti, vector RAG, or PostgreSQL?
Compare Graphiti, vector RAG, and PostgreSQL for AI agent memory. Separate truth, retrieval, history, and verification before production decisions at scale.
Validating tool calls in TypeScript for AI agents
Validate AI agent tool calls in TypeScript with Zod, structured errors, safe output, and bounded retries before reaching your API.
Software engineering skills for coding agents
Learn which software engineering skills remain valuable with coding agents: specification, architecture, context, review, evals, and operations for real teams.
How to Set Up AI Code Review in GitHub Actions
Set up AI code review in GitHub Actions with Codex, least privilege, and separate jobs. Includes YAML, verification, errors, and safety limits.
How to Build a TypeScript MCP Server for Coding Agents
Build a TypeScript MCP server, test it in Inspector, and connect it to Claude Code and Codex with schemas, output limits, and read-only tools.
The agent stayed sandboxed; the host trusted the file
In 2026, Pillar published 7 escapes across 4 agents; protect the host, Git, IDEs, hooks, Docker and CI from trusted workspace files.
When an agent writes a migration, the database should say no first
In 2026, 84% of developers use or plan to use AI; review agent-made database migrations in CI with dry-runs, lint, and rollback before real data is touched.
Without telemetry, a coding agent becomes a black box
In 2026, GitHub reported 60 million Copilot reviews; build coding agent observability in CI with JSONL, OTel, safe proof, cost control, and reviewable evidence.
PR comments are not trusted context for agents
In 2026, GitInject documented 11 attacks across 4 providers; protect GitHub Actions agents from PR prompt injection with gates, sandboxing, and minimal secrets.
The agent patch passed; the product still broke
In 2026, TDAD cut test regressions by 70%; build regression evals for coding agents with impact analysis, CI, and PR proof.
Codebase RAG only works when the agent asks well
In 2025, Cursor measured 12.5% higher accuracy with semantic search; build codebase RAG with grep, symbols, MCP, and PR proof.
Agent PRs die when they become reviewer labor
In 2026, a study analyzed 33,596 agentic GitHub PRs; reduce reviewer abandonment, duplicate work, large scope, and broken CI before safe human review.
When EDR mistakes an agent for an intruder
In 2026, Sophos saw credential access in 56.2% of agent-related blocks; tune EDR, permissions, and CI without opening secrets.
The hidden context cost in coding agents
In 2026, Systima measured about 33k tokens before the prompt in Claude Code; build a context budget for MCP, instructions and subagents with CI proof.
Friendly Fire turns the security agent into the exploit
In 2026, AI Now showed RCE in Claude Code and Codex during third-party repo review; isolate permissions, sandbox, network and CI evidence before risky PRs.
Your agent can clone a repo that never existed
In 2026, arXiv measured up to 85% hallucination in repo cloning; lock origin, MCP and CI before the agent executes.
GhostApproval shows when the agent diff lies
In 2026, Wiz tested 6 coding assistants; learn how to review symlinks, sandboxing, and approvals before agents write outside the workspace.
Unknown repos need an agent quarantine step
In 2026, 0DIN showed a payload outside the repo; use agent quarantine with sandboxing, denied network, and no setup secrets.
MCP allowlists keep coding agents on rails
In 2026, MCP research found 57 threats; use tool allowlists, private registries, CI gates, and logs for coding agents.
When CI sends the failure back to the agent
Build a self-correcting agent loop in CI with short logs, sandbox, retry limits, and reviewable PR proof without opening broad diffs or losing control.
Executable specs make coding agents obey
In 2026, GitLab found 85% see review as the bottleneck; use executable specs so coding agents prove each change.
Sentry agentjacking changes the risk model for coding agents
In 2026, Tenet reported 2,388 exposed organizations; treat Sentry MCP events as hostile input before Codex or Claude Code acts.
Hooks stop coding agents before the damage
Claude Code documents 3 hook cadences; use gates before, after and at the end of the loop to reduce real risk across coding agents, MCP, shell and CI.
A lean AGENTS.md helps agents spend context better
In 2026, one study measured over 20% higher cost from weak context files; write a lean AGENTS.md so Codex and Claude Code test better without token bloat.
Subagent fan-out for large code migrations
In 2026, 85% see review and validation as the bottleneck; use subagent fan-out to migrate monorepos with proof, worktrees, and lean PR evidence for review.
Codebase RAG over MCP for agents that should not read everything
In 2026, more than 1 in 5 GitHub reviews already run through Copilot; use codebase RAG over MCP for agent context on demand.
PR evals that keep code agents honest in CI
In 2026, GitHub had seen 60 million Copilot reviews; build PR evals in CI to validate agents before the reviewer.
Context engineering for coding agents without waste
In 2025, DORA measured 90% AI adoption in software teams; use lean context so coding agents work with better evidence.
The Bottleneck for Code Agents Is Review
In 2026, 85% of devs cite reviewing, editing, and testing AI code as the bottleneck. See a practical harness for reliable PRs, with gates, MCP, and subagents.
Claudiomiro: How to Achieve 100% Autonomous (Complex) Coding
Claudiomiro is an AI-powered CLI that autonomously runs the full software development lifecycle: analyzing codebases, planning, and implementing complete…
Creating a Native Knowledge Graph System with MongoDB, S3 and LangChain
A practical guide to building a native knowledge-graph system with MongoDB, S3 and LangChain that links entities and relationships to give AI real context.
Building a Fitness AI Agent with LangChain.js: A Real-World Implementation
Deep dive into implementing a production-ready AI fitness agent using LangChain.js with structured tools, proper prompting, and intelligent context retrieval.
Mastering Langchainjs to Coordinate Tool and Function Calls Using AI Agents
Discover how Langchainjs orchestrates AI tool and function calling with the chain-of-thought paradigm, featuring a practical fitness app example.
Long-Term Memory for AI: How Graphiti Works for Building Real Smart Applications
A deep dive into Graphiti, Zep's open-source temporal knowledge-graph framework, plus a full blueprint for building a Personal Financial AI Coach that learns…
Why I Hired ChatGPT Pro ($200/Month) as My Junior Developer
Discover how hiring ChatGPT Pro as a junior developer at $200/month revolutionized my software engineering workflow, boosting productivity and efficiency.
