Kamila’s latest post on the Substack really hit home regarding the 'agentic harness' sprawl. I’ve been actively trying to avoid this complexity tax on my personal aggregator project, Daily Relay, which runs completely local on a Node Alpine container.
Instead of letting an autonomous agent self-route or handle control flow, I built an explicitly bifurcated harness. Hard-coded JS handles task lists and calendars with 0ms latency, while a local 3B model is tightly boxed into a single-item parsing path. To make the 3B model production-grade for TTS, I built a lightweight evaluator script using regex score deductions for forbidden formatting alongside a local LLM judge. Ended up cutting generation latency by 56% just by tuning the constraints.
Thank you - if not’s not harness noise, it’s now loops noise, both of which make it seem like utility requires some unachievably high complexity of systems… Not necessarily
Kamila’s latest post on the Substack really hit home regarding the 'agentic harness' sprawl. I’ve been actively trying to avoid this complexity tax on my personal aggregator project, Daily Relay, which runs completely local on a Node Alpine container.
Instead of letting an autonomous agent self-route or handle control flow, I built an explicitly bifurcated harness. Hard-coded JS handles task lists and calendars with 0ms latency, while a local 3B model is tightly boxed into a single-item parsing path. To make the 3B model production-grade for TTS, I built a lightweight evaluator script using regex score deductions for forbidden formatting alongside a local LLM judge. Ended up cutting generation latency by 56% just by tuning the constraints.
Thank you - if not’s not harness noise, it’s now loops noise, both of which make it seem like utility requires some unachievably high complexity of systems… Not necessarily