CURRENT / Sources & referenceSource status

Current

Opening the reference…

Latest

What’s new. What’s useful.

New discoveries, practical guides, and changes worth knowing.

Browse the archive →

DiscoveryOct 16

technique

EvalSkills

Eval Skills gives coding assistants procedures for inspecting AI executions, collecting human error annotations, building graders and comparing changes. Start with one real application trace and a person who can judge the expected behavior. The workflow then turns observed failures into reusable test cases and experiments. This is useful for improving a particular AI feature, not assigning every model a universal quality score. Judge-provider setup is a separate choice. Current reviewed the README; it has not installed the suite or validated a resulting grader.

DiscoveryOct 16

product

FLUX.3 image Klein

FLUX.3 Image Klein is a reference-image editing model with caption and layout controls. Try it for a color variant that must fit an existing character rig, checking both silhouettes and attachment points. Boxes guide rather than mask the edit; gated access, terms and hardware need checking before use.

DiscoveryOct 16

workflow

GameFactory-3A

GameFactory-3A is a collection of agent skills, asset pipelines and engine adapters for building game prototypes. Its documented path starts with a coding agent reading the project’s setup overview, then selecting the engine and asset work. The demos combine generated objects with downloaded or existing assets; they do not establish that an entire commercial game was produced in one step. This can help developers explore an asset-to-engine workflow, but model providers, hardware and asset permissions depend on the selected path. Current has not executed a demo.

DiscoveryOct 16

feature

n8n Agents alongside workflows

n8n Agents adds reusable, conversational agents alongside existing workflows. You choose instructions, a model and permitted tools; an agent can call a workflow, while a workflow can call a published agent. This suits variable tasks such as investigating a support request while keeping specific actions inside fixed workflows. The September 25 announcement describes sessions, versions and sensitive-tool approvals, with the product still in preview. Agent turns share the execution quota, and model access needs credentials or Gateway credits. Existing AI Agent nodes remain supported.

DiscoveryOct 16

creative-use

Underbloom reference-first creature production

Underbloom’s creator describes building creatures from approved references, cutting separate image parts and fitting them to procedural rigs. The useful pattern is to test one assembled character before generating variants. The companion exercise uses an original four-part courier; Current has not play-tested the game.

DiscoveryOct 16

product

BeelineBench decision-model evaluation

BeelineBench evaluates how efficiently decision models navigate multi-step problems by comparing explored states with a shortest path. It includes heuristic, random and oracle baselines across seven domains. This is a planning benchmark, not a ranking of general assistants or business-triage quality. Its search has a 255-state frontier limit and a 2,500-node stopping limit; provider responses can vary even with fixed trial seeds. The model-run command makes paid requests. Current read the method and limitations but has not reproduced the results.

DiscoveryOct 16

product

Blokkily

Blokkily is an experimental music workstation with synchronized pattern editors and an arrangement timeline. An AI composition assistant proposes musical changes for review, and an MCP bridge allows agent control. The useful distinction is between editing an inspectable song and accepting a generated audio file: the arrangement remains available for revision. Its documented export renders the arrangement, so a live pattern performance must first be recorded there. Plugin and operating-system requirements vary. Current checked the README, but has not installed it, listened to exports or tested the assistant.

DiscoveryOct 16

product

Poppy Personal Agent Protocol draft

Poppy is a draft Personal Agent Protocol for assistants working with businesses through discovery, sign-in and shared sessions. Use its concepts to specify an explicit reservation-change handoff and confirmation receipt. Instinct is a named participant; that does not establish a working integration in every account.

DiscoveryOct 16

product

OpenRig

OpenRig organizes multiple coding-agent sessions into persistent teams with roles, shared context and assigned work. Its setup supports Node.js 22 or 24 with tmux on macOS or Linux, with a WSL2 route for Windows; the README recommends Node 22 on Apple silicon. This can help coordinate a repository change across specialists. Launching a rig changes provider hooks and workspace trust settings, so inspect its dry-run and configuration guidance first. Current reviewed the setup documentation but has not run a rig or tested cross-agent isolation.

DiscoveryOct 16

product

Golden Thread vault memory and scoped rule checks

Golden Thread keeps project memory in an Obsidian vault for Claude Code sessions, with procedures for recording decisions, resuming work and handing off unfinished tasks. It also uses hooks to check selected rules. The README now identifies plugin v0.21.0, correcting our earlier v0.20.5 reference, and distinguishes validated rules from reminders. It may help teams repeatedly rebuilding project context. Its token-saving results are author-reported, and the hooks do not stop an already-running process with your privileges. Current has not installed or independently audited it.

DiscoveryOct 16

research

Assuring AI-written software

This October 6 case study examines a healthcare platform built with coding agents and supervised by an operator without formal software-engineering training. It reports failures in the oversight itself: checks that did not run, monitors that measured the wrong outcome, and a disruptive automatic repair. It is useful when designing agent supervision because a clean summary can conceal missing work. Our related workflow guide demonstrates an omission check. One reported case does not establish a general reliability or safety rate.

DiscoveryOct 16

research

EvalResearchBench

EvalResearchBench studies whether agents can build useful evaluation suites under fixed time and API budgets. Researchers create tasks and graders, freeze the evaluator, and compare its model ordering with target benchmarks, including targets hidden during development. The practical lesson is to judge an evaluator on separate examples, not only the cases used to improve it. The October 3 paper also reports truncation and budget failures. Current checked the abstract; its experiments and headline scores have not been reproduced.