AI Agent Struggles to Master Basic Strategy in Slay the Spire Training Run

·3 views

A developer trained a reinforcement learning agent for 200,000 steps — roughly 14,000 fights — against the Cultist, the simplest enemy in Slay the Spire. Despite extensive training, the agent achieved only a 70% win rate, effectively playing cards at random rather than learning a deliberate strategy. The agent also consistently refused to use block cards, even when energy and cards were available after attacking. Investigations revealed a silent bug had disabled block mechanics, but fixing it did not change the agent's behavior, pointing to a deeper issue with delayed rewards in reinforcement learning. Various reward-shaping experiments and hand-coded heuristics improved HP preservation somewhat, but gains repeatedly degraded during further training.

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)

DEV Community's Top 7 Posts: AI Hiring Shifts, Goroutines, and Prompt Injection

The DEV editorial team selected seven standout community posts from the week of August 22–28, covering topics ranging from tech hiring trends to AI productivity and software engineering experiments. A developer shared observations from two months of job hunting, noting that employers in 2026 increasingly favor exact-skill matches over candidate potential. Other featured posts explored why structural design beats prompt-based safety in AI agent engines, and what developers should do during the idle gaps AI coding tools create. A performance deep-dive revealed how spawning a million Go goroutines can balloon RAM usage far beyond the commonly cited 2KB-per-goroutine estimate. Rounding out the list were creative pieces comparing hiring processes to HTTP status codes and a real-world account of the bugs that emerge when flat JSON files replace a database.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Developer Connects Instagram Fake-Follower Analyzer to Slack Using Apify MCP Connector

A developer integrated their Instagram Influencer Deep Analyzer — an Apify Actor that scores accounts for fake-follower probability — with Slack to enable real-time alerts during batch runs. Previously, suspicious accounts could only be identified after an entire scrape job finished, sometimes hours later. Using Apify's MCP connector as a proxy, the Actor forwards high-risk alerts to Slack mid-run without the code ever directly handling Slack credentials. During a 20-profile test batch, the system flagged an account with 78% fake-follower probability — citing a skewed follower ratio and sub-0.3% engagement rate — while the Actor was still processing the next profile. A key technical challenge was avoiding hardcoded Slack tool names, which the developer solved by dynamically detecting the correct messaging tool and argument schema at runtime.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

GitHub Actions Pipeline Uses AI to Draft Docs While Keeping Humans in the Loop

A developer has shared a documentation review workflow that uses AI to generate draft documentation from code diffs on every merged pull request touching source files. The pipeline, built as a GitHub Actions workflow, sends code changes to a language model that produces structured sections covering what changed, usage impact, and items flagged for human decision. The system deliberately splits responsibilities: the model handles parameter descriptions and usage examples, while humans are required to verify security implications, business logic, and deprecation notices. A review issue with a checklist is automatically opened after each merge, ensuring that AI-generated content is formally signed off before being treated as accurate. The author notes that stale or unverified AI-written comments pose a real documentation risk, and that this checkpoint-based approach addresses that without abandoning AI assistance altogether.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Tutorial: Run LLMs Locally on Mobile Using React Native and Expo

A new developer tutorial demonstrates how to run large language models directly on a user's device using React Native and Expo, requiring no server or API key. The guide uses the open-source NobodyWho library, which wraps llama.cpp in Rust and exposes a React Native API for running GGUF-format models locally. Models can be automatically downloaded from Hugging Face, cached on first use, and reloaded on subsequent app launches without manual management. On-device AI offers key advantages over cloud-based approaches, including offline functionality, user privacy, lower latency, and zero inference costs. The tutorial covers a range of features from basic chat to multimodal input, speech-to-text, voice detection, tool calling, and retrieval-augmented generation.

0 comments Read more at DEV Community