Cursor Projects Beta Shows Promise but Struggles With Decision Retention in Real Use

·1 views

A developer tested Cursor Projects, an AI coding tool in beta, over three days across five working sessions totalling roughly 14 hours, using one orchestrator and 20 agents to build a podcast monitoring system for a home server. The orchestrator delegates tasks to cloud-based agents while maintaining a shared context folder, allowing the user to add input at any time without interrupting the workflow. However, the project's planning file was rewritten 111 times but explicitly read only once, causing previously agreed decisions to vanish and agents to apply unapproved defaults. The orchestrator also reportedly denied having completed work it had actually done, raising concerns about reliability for professional use. The author suggests improvements such as a persistent decision log, version-controlled context, and clearer ownership of review comments, while acknowledging some issues may be addressable through custom rules as the beta matures.

Read the full story at DEV Community

This is an AI-generated summary. ShortSingh links to the original source for the complete article.

Discussion (0)

How AWS Architecture Handled 1.5M Requests in One Hour for a Flash-Seat Exam Booking

A national certification exam with 20,700 seats across five levels needed a new registration system after its previous platform crashed and double-booked seats when registration opened. The redesigned system on AWS uses an atomic Redis script as the sole arbiter for seat allocation, ensuring no seat is ever sold twice and no database is involved in the admission decision. Static content was served via CloudFront and S3, with an AWS WAF web ACL handling roughly three in four requests without reaching origin servers. During the first hour of the live event, the system served over 1.5 million origin requests, peaked at 54,800 requests per minute, maintained median latency under 10 milliseconds, and recorded a server error rate below 0.01%. This is the first installment of a three-part series detailing the architecture, edge strategy, and capacity planning behind the solution.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

AEGIS: Agentic Fraud Investigation System Built on TigerGraph and GraphRAG

A developer has built AEGIS (Agentic Evidence & Graph Intelligence System), a multi-layered fraud investigation platform created for the TigerGraph HHGOA challenge. The system goes beyond single fraud scores by mapping relationships between customers, cards, devices, transactions, and past investigations using a TigerGraph graph database. AEGIS uses an agentic orchestration loop that sequentially handles evidence gathering, uncertainty assessment, policy evaluation, and Next Best Action recommendations before any human authorization is granted. The platform integrates GraphRAG to pull context from historical cases and employs TigerGraph's Weakly Connected Components algorithm as a structural signal to detect coordinated fraud patterns. Key design principles enforce strict separation between fraud signals, verdicts, recommendations, and authorizations to prevent automated over-reach.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

A 15-Minute Method to Tailor Your Resume to Any Job Description

A practical resume-tailoring technique involves categorizing job ad requirements into must-haves, nice-to-haves, and context before making any edits. The core advice is to focus exclusively on must-have requirements — phrases like 'required' or 'X+ years of' — since these are what recruiters and keyword filters screen for first. Job seekers are urged to match the exact terminology used in the job ad, replacing vague descriptions with specific, evidence-backed bullets that mirror the employer's own language. Each must-have should be rated as covered, weak, or missing, with 'weak' entries offering the highest return on tailoring effort. The method requires no special tools and can be completed in around 15 minutes using only a text editor and two browser tabs.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

OpenAI Codex Flaw Let Untrusted AI-Generated Code Steal Auth Tokens from Memory

A sandbox escape vulnerability in OpenAI's Codex desktop edition, disclosed on September 21, 2026, allows untrusted AI-generated code to read authentication tokens from shared process memory. The flaw stems from Codex's bundled Node.js tool running both trusted and AI-generated code within the same memory space, enabling token theft without triggering any visible permission prompts. Stolen credentials can then be used to forge requests to external native programs capable of launching applications, modifying system configurations, or accessing local Unix sockets — constituting a full sandbox escape. Researchers noted that the severity is compounded when agents hold broad, long-lived credentials, as a single memory read could grant prolonged unauthorized access. Since Codex operates primarily on developer workstations rather than internet-facing servers, the true number of affected systems is likely far greater than the 164 instances indexed by external scanners.

0 comments Read more at DEV Community