Agents In UseA closer look at the steps, decisions, and human checkpoints behind agent work.
Liam builds an n8n Telegram assistant connected to Todoist, Gmail and an expense spreadsheet. The recording keeps several useful failures: an OpenRouter privacy-policy mismatch, an image-reading problem and a receipt whose handwritten tip is missed. A manually triggered schedule creates tasks, and a follow-up instruction adds links to the task messages.
Bart connects Hermes to a local model through LM Studio, adds Honcho memory and Brave Search, and sets up Telegram delivery for a daily AI-news task. The small model needs a reminder that it is already inside Telegram before creating the schedule. The walkthrough is useful about dependencies, but its cloud memory and search services mean the complete setup is not fully offline.
A Hermes quickstart connects a locally installed agent to Telegram and iMessage, then creates a diagram with a skill and schedules a sports-and-AI-news message. The transcript is useful for showing the gateway dependency and the allowed-user setup. It also distinguishes an agent replying in a terminal from a messaging integration actually reaching a phone.
Video listing based on its title and description; the recording has not been reviewed.
Video listing based on its title and description; the recording has not been reviewed.
Video listing based on its title and description; the recording has not been reviewed.
Video listing based on its title and description; the recording has not been reviewed.
Video listing based on its title and description; the recording has not been reviewed.
Video listing based on its title and description; the recording has not been reviewed.
Video listing based on its title and description; the recording has not been reviewed.
Dazza Greenwood evaluates a sample agent trace in Lake Merritt using JSON in an OpenTelemetry-style format. The hypothetical user asks for a nearby coffee shop and a table booking. The judge gives partial credit because the agent finds the shop but cannot book a table. The important lesson is to score the whole requested outcome, not the plausibility of the final reply.
Video listing based on its title and description; the recording has not been reviewed.
This Lake Merritt walkthrough moves evaluation instructions into a versioned YAML eval pack, then applies the pack to a CSV of sample answers. The presenter inspects a Hamlet summary that scores poorly and questions the expected answer itself. It is a useful reminder that an evaluation can encode the wrong preference even when the machinery runs correctly.
Video listing based on its title and description; the recording has not been reviewed.
Video listing based on its title and description; the recording has not been reviewed.
Joe configures a Codex automation to make a daily business note from an Obsidian vault, email, calendar and transcripts. A pre-existing Markdown template supplies the structure. The recording includes an emoji-path problem and a review of the schedule, timezone and selected model. Crucially, the presenter does not run the final automation on camera.
The presenter prepares for content-marketing interviews with ChatGPT deep research and an ElevenLabs voice agent. Company research, the job description, a résumé and past results inform likely questions. A voice agent then asks those questions and follows up on answers. The presenter reports two offers while also explaining the substantial human research and editing involved.
Video listing based on its title and description; the recording has not been reviewed.
Dazza Greenwood adapts the BBQ fairness benchmark to Lake Merritt with a custom scorer and eval pack. The demonstrated run uses an age-related subset and compares predicted labels with expected labels. The workflow illustrates why reproducing a benchmark needs its metadata and scoring rules as well as a list of questions.
Video listing based on its title and description; the recording has not been reviewed.
Video listing based on its title and description; the recording has not been reviewed.
An AirOps workflow turns a Fireflies transcript link already arriving in Slack into a Google Doc containing an outline and action items. The presenter runs it on a podcast interview and inspects the generated document. This example is specific about the missing connection: retrieve the transcript text behind the link before asking an LLM to organise it.
Video listing based on its title and description; the recording has not been reviewed.
Video listing based on its title and description; the recording has not been reviewed.
