Real-world agent-assisted engineering at scale — adoption, governance, security risks
Key Questions
What tools are used in real-world agent-assisted engineering?
Key tools include Cursor3, Claude Code, ATLAS, LangGraph, Auton, Pydantic, and n8n. These enable scaled adoption.
What are some success stories of agent deployment?
Amex achieved 30% efficiency gains, with wins at Pinterest, Starling, and Block's Managerbot. Managerbot is a proactive Square AI agent proving Jack Dorsey’s AI bet.
What is GLM-5.1's performance in benchmarks?
GLM-5.1 is SOTA on SWE-Bench, topping open-source and ranking #3 globally, beating GPT-5.4 and Claude Opus 4.6.
What pitfalls exist in agent-assisted engineering?
Pitfalls include deskilling, errors, hallucinations, and agent vulnerabilities, as studied by Stanford, DARPA, and reports on autonomous exploits. Multi-agent setups don't always improve results.
What security risks do AI agents pose?
Every deployed AI agent can be turned against you, requiring strong incident response. Governance, guardrails, and observability are essential for production.
Should companies build or buy agent solutions?
Real deployments highlight build-vs-buy decisions, with lessons on governance, evaluation, and risks like hallucinations in LLMs.
How is agentic AI transforming software engineering?
AI boosts productivity but predicts disasters in usage; coding models reshape roles, per Simon Willison.
What alliances support enterprise AI transformation?
McKinsey and Wonderful AI teamed up for agentic AI delivery from ambition to production.
Cursor3/Claude Code/ATLAS/LangGraph/Auton/Pydantic/n8n; wins (Amex 30%/Pinterest/Starling/Block Managerbot); GLM-5.1 SOTA SWE-Bench; pitfalls deskilling/errors/hallucinations/agent vulns (Stanford/DARPA/autonomous exploits); build-vs-buy.