35,684 results · Most relevant
To run agents in a Mac or other non-hosted environment, define a custom tool for your agent as part of the workflow run; when the custom tool call comes in, do whatever you want on that machine and return the result back to the model. You can also run the environment self-hosted in your own infrastructure and bring your own sandboxes.
To start a code layer from an existing app, clone the repository (e.g., 'Pickup Marketplace' on the main branch) rather than prompting from scratch.
Select the code layer before typing a prompt so the agent operates on the right layer.
Listen to your audience about what they actually want, then make products people are already asking for; trying to trust your own best judgment on what products to make is the inverse of listening to your audience, and if there is enough volume on audience product requests, just make the product.
For hard conversations about underperformers, the speaker recommends the frame: 'I'm going to take on this guilt and pain so that everyone else here doesn't have to suffer.' He calls taking on the guilt of hard people decisions one of the more charitable frames for making them.
Randy Hunt argues for keeping design teams small and dense with strong people: small size reduces communication overhead, constrained bandwidth becomes a forcing function for prioritization, and a team that is too small becomes a delivery-only production engine, which he calls a fail state.
Harish Gaggar describes an onboarding flow: register the agent in a system with its use case and vertical to generate an agent ID, submit details (framework, language, skills/prompts) through an internal dev portal, then clone the generated repo, build a Docker image, and test locally before deployment.
Harish Gaggar describes sending all operation logs from agent calls to a log store (e.g., ELK or a cloud virtual server) the moment calls are submitted, sending real-time traces and metrics to OpenTelemetry, and saving aggregated data into Grafana for visualization; the log store can answer operational questions like 'show me all 401s in my last 10 minutes of calls'.
Harish Gaggar describes the agent's query handling: interpret the question and create a plan, classify it as read, modification, deletion, or sensitive operation, fetch evidence via the authenticated MCP gateway and MCP server, then either execute read-only operations (logging to audit logs) or route write/sensitive operations through human-in-the-loop approval.
Harish Gaggar warns that building only an agent without the underlying enterprise frameworks leads teams to retrofit those frameworks after problems occur; he recommends putting the frameworks in place from the very first place.
Harish Gaggar describes implementing agent-level tracking and eval separately from gateway monitoring: each agent gets its own project and tables (job tracking, quality metrics, eval scores, tool invocations, agent trajectory, etc.), with eight quality checks per run scored 0–1, a nested quality deep-dive row, tool invocation records, and a step-by-step trajectory of LLM calls and graph state.
Harish Gaggar describes implementing agent memory management to save and restore workflow state over time: deeper questions execute on a particular LangGraph node and save a checkpoint there, follow-up questions create a new run ID and save checkpoints on their nodes, with SQLite for local development and Cloud Spanner (GCP) or equivalent databases in Azure/AWS for production.
Logging must be sufficient to catch an incident while it is happening, with real-time alerting rather than data buried in logs — this reduces detection from months to minutes.
A code freeze communicated only as a chat message is not a real constraint; it is just asking nicely, unless there is an infrastructure-level restriction. When prompting a coding agent to investigate issues, tell the agent to consider the command an active code freeze and only investigate without changing anything. And when an agent encounters ambiguous results like empty query output, it should ask a human what that means and what it should do before taking destructive action.
For an AI deployment, want a concrete account of the intended use, a threat model, the failure modes, and a response plan — the same questions Paden Gayle needed answered at Google: what happens when the system breaks, who notices, what stops, who takes responsibility, and how long to bring it back up.
For a workflow, separate clean documents, scans, rare exceptions, and new document types; measure errors and control behaviors in each segment group, then lead with the weakest result. An average can hide a serious gap in one segment; if a group has little evidence, say so and restrict that use until you have enough.
Generate the evidence with each evaluation run; alongside it, preserve the input, prompts, model version, configuration, output, and reviewer action. Calling the model again later may produce a different answer, so preserving what actually happened lets you reconstruct the historical decision.
In the closing story, the demo worked and the model was fine, but it still did not ship. When a working demo does not ship, go find the people. Treating AI deployment in a bank as purely an engineering problem causes the project to stall despite a working demo.
Enforce policy rules beyond CI — just because code compiles doesn't mean it follows specifications like HIPAA or DORA. Agents won't be aware of compliance policies unless your systems enforce them on the other end, and you should record that enforcement happened for auditability.
Dr. Andy Galpin recommends keeping a consistent weekly structure and making small exercise variations — such as close grip bench press on bench day, slightly changing range of motion, or using a special type of barbell — as in the conjugate model used by powerlifters, rather than swapping exercises entirely.
Choose sources → to get a library of your own and a daily brief.