Top AI Demos #24: Deterministic agent pipelines, code leaks and agent memory

Jason Gardner presented hiveWiki.ai, a wiki-based collaboration layer where independent agents, Claude Chat, a Cursor Agent, and a human share the same usable “wiki artifact” for transparent coordination via MCP. The demo showed how agents could draft, critique, and deploy a feature while keeping the context consistent and reviewable. For anyone building toward more autonomous workflows without the brittleness, this felt like a practical path to agent memory and human-in-the-loop clarity people kept coming back to.

Justin Bergeron demoed his Mac Mini AI setup that worked like a scrappy local “dev team” for HausHavn, splitting responsibilities across architecture, coding, QA, and PM, then orchestrating handoffs into a coherent software workflow. The project emphasizes role design over one-shot prompting, using local agent coordination and practical automation to keep tasks moving. It stood out because it tackled the messy reality of making agents useful as process infrastructure, not magic, and (people kept saying this felt immediately adaptable). We liked how it points toward agentic workflows that can scale cost-effectively on modest hardware.

Bob Prendergast showcased a real-time MoE neural telemetry engine that visualizes Mixture of Experts routing decisions from IBM Granite 3.0. The demo used a PyTorch forward hook to capture 40-dimensional routing weights from Layer 20 and render a live laptop heatmap of expert activation. After that, the model reflected on what it was doing by interpreting its own routing pattern, and people seemed to really enjoy that loop. For builders chasing efficient, local inference and agentic introspection, it felt like a practical blueprint for turning sparse activation into something you can actually debug, demo, and productize.

Alula Zeryihun presented a multi-agent sales engine that kills guesswork in outbound by running seven specialized agents before any outreach happens: identity, profile, activity, market, network, scoring, then synthesis. The workflow grounds decisions in live web data, passing structured JSON contracts between stages to keep messaging context-aware while controlling cost and latency. It felt especially relevant to builders because it turned a “single-shot prompt” into a decision pipeline, and people seemed to love the clarity of the orchestration. The demo also fits the current shift toward more agentic, local-first automation for regulated, real-world work.

Henry George from Super44 presented Boomerang, a custom agentic browser infra built as a replacement for tools like browserbase and Firecrawl, focused on reliable web automation. The system handles stealth-style browsing and orchestrates agent web flows while working through messy realities like Cloudflare friction, so the agents fail less and recover better. Given his hands-on serverless, event-driven background, it felt like an infrastructure case study you can actually reuse. We liked it because builders (people seemed to) want practical guidance on agentic web usage, right when high-speed local inference and long-horizon autonomy make this layer increasingly necessary.

Felipe Pacheco from Poweredia presented Anthropic Ambassador Chile, a local-first push to help Latin America move from AI consumers to AI builders. The demo brought together Claude community leadership, Bendita IA’s hands-on builder ecosystem, and his Agent OS approach to orchestrating intelligent agents with clear organization-level governance. It tied fintech experimentation to practical tooling, and it landed well with the community, with feedback hinting that people wanted more “how to build” paths. We liked it because it turns agentic ambition into repeatable developer momentum.

J.D Nicholls spotlighted “Ingeniería Inversa de Plugins de Claude Code” and showed how the repo notebooklm-ai-plugin turns NotebookLM into a terminal-first workflow by reverse engineering the free RPC path using browser cookies plus Chrome CDP. He implemented the plugin in TypeScript and Bun, packaging multiple artifact types with a single .claude-plugin manifest for easy distribution. We liked how people leaned in to try it immediately (people loved it). It fit the broader agentic tools trend: practical orchestration you can actually run, adapt, and ship.

Thang Chung from Tiger Tribe presented a demo on adaptive tool orchestration with on-demand discovery and code mode execution. The project runs an MCP-based transform that does contextual tool resolution, then generates controlled C# workflows using the GitHub Copilot C#/.NET SDK and the MCP .NET SDK to chain calls with less prompt bloat. We liked how it shifted “pick from everything” into “query, narrow scope, then execute safely,” and the audience feedback quietly leaned positive. It also points toward a more product-ready, enterprise-friendly agent pattern where tool catalogs stay lean while orchestration stays deterministic.

Alain Krok from AWS shared an autonomous background coding agent demo that runs long horizon tasks without constant supervision. The agent kept working in the background, managing state and task progression while tackling multi-step coding changes like an “always on” engineer. Given his systems plus full-stack experience in C++ and cloud tooling, he focused on the real engineering friction: reliability, focus, and safe continuation across long runs. People also seemed to like how pragmatic it felt for builders staring at codebase-scale work, and it pointed to the next shift toward dependable agentic execution.

Juan Campos Gomez presentó Pista Inteligente: un pipeline end-to-end que convierte PDFs y texto de carreras en CSV con Claude como ETL estructurado, luego entrena un LightGBM LambdaRank con Walk-Forward Validation para ordenar por probabilidad de victoria y calibra con isotonic regression para que el score sea interpretable. Todo corrió en Python con mlflow para tracking de experimentos. Lo que gustó (people loved it) fue lo reutilizable de tres ideas: prompts con header fijo para colapsar ambigüedad, ranking vs regresión y calibración sin reentrenar. Además, el caso real para datos hípicos en Chile lo vuelve cercano y con potencial de producto tipo “predicciones con explicación” para apuestas o gestión.