Know Today

AI agents Developer workflows Reusable skills Model routing Verification loops Human review Context isolation

👀 Keep an eye on

  • Reusable agent skills can turn requirements intake, project setup, and recurring checks into versioned workflow components.
  • Per-task model and context routing makes parallel agent work more affordable when routine checks do not need a full project history.
  • Reliable automation increasingly means pairing execution with explicit pass/fail verification, run history, and bounded retries.
  • Separate producer and reviewer agents can reduce correlated mistakes when both work from clear acceptance criteria.
  • Human approval gates remain essential before publishing, spending money, changing production systems, or taking other irreversible actions.

🛠️ Practical workflows

PatternUseful implementation
Scripts plus judgment Move stable fetch, transform, and validation steps into tested scripts; leave interpretation and exception handling to the LLM.
Context boundaries Use separate sessions and folders for unrelated work, then give each agent only the files and criteria it needs.
Skill hardening Record recurring edge cases and failed assumptions in the workflow itself, then review that list as the process evolves.
Structured planning Have the agent ask focused questions before work begins, especially around constraints, success criteria, and ownership.

💭 Opinions worth testing

Opinion: The useful unit of agent work is a loop, not a prompt: execute, verify, record the result, then retry only within clear limits.

This should influence whether a team invests first in acceptance criteria and observability rather than collecting ever-larger prompt templates.

Opinion: “Middle-to-middle” workflows—human framing, agent execution, human validation—are more dependable than fully autonomous end-to-end operation.

This favors designing review checkpoints around costly downstream errors instead of removing people from the process.

⚠️ Caveats

  • Needs verification: Reported speed gains, cost savings, and product capabilities are anecdotal or may change; confirm current behavior and policy before standardising a workflow.
  • Risk control: Auto-approval, broad tool access, retries, and concurrent agents can compound cost and impact, so start with approval pauses and spending limits.
  • Practical limit: An execution loop is only as trustworthy as its verifier; vague success criteria merely automate ambiguity.

✨ Try this today

1. Build one bounded project-start skill

Write a short reusable instruction that asks five requirements questions, creates a local checklist, and stops for approval. Expected outcome: less repeated setup prompting and a clearer handoff into implementation.

2. Split a small task into producer and reviewer passes

Ask one LLM session to draft a change plan and another fresh session to review it against three acceptance criteria. Expected outcome: gaps become visible before you touch files or run commands.

3. Extract a deterministic step

Choose a repeated task such as formatting, test selection, or file inventory, and replace the conversational instruction with a tested local command. Expected outcome: faster, more repeatable agent runs with less context spent on mechanics.

4. Add a verifier before a retry

For one laptop-scale automation, define a concrete check such as “tests pass and the expected file exists,” log the result, and allow only one manual retry. Expected outcome: you learn whether the workflow is measurable before increasing its autonomy.