Accidental Tech Podcast680: A Lot of Holes in That Cheese
https://atp.fm/680
Our home screens, desktop audio setups, document-saving models, what we hope to “experience” next week, and where not to aim snow.
Nathan King@atpfm @siracusa I’m a huge proponent of your work, but your view on AI is a rooted in a standard chat interface. Agent orchestration is here, with teams of agents with specialized skills swarming on an issue. One assistant can make dumb mistakes, but a *team* with a reviewer recognizes issues and iterates until all issues are resolved. Be prepared for your relationship with coding to change this year.
Source: I’m a senior AI engineer working at a midsize bank.
Marco Barrios@NathanKing, what do you mean by specialized skills? Models trained for specific programming-related purposes?
@marcobarrios You can create skill files relevant to the specific strategy, development practices, or architecture of the project. They do not pollute the context window because only the YAML frontmatter is exposed to the agent. Only if the task is relevant will the agent load the skill information. This can be very powerful when chained together and combined with tool use. For example, coder > reviewer skill using linting/Playwright > feedback back to coder in a loop.
https://platform.claude.com/docs/en/agents-and-tools/agent-skills/overview
@marcobarrios That’s not to say the code won’t still have semantic errors, but that’s what good test driven development will catch. You can further extend this with an orchestrator that will parallelize work when possible (and split between agents to avoid context bloat). Now you can start to see a path toward genuinely good quality software delivered very quickly by a small team.