@gptproto-jacob.bsky.socialNot every failing agent needs fine-tuning.
Rewriting the harness lifted Qwen3.5-4B
from 0.16 to 0.30 on held-out tasks,
but never cut the share of bad plans.
A LoRA cut those from 28% to 5%.
Loops are a harness bug; bad plans are a
weights bug. #AgentHarness
@mollzmp.bsky.socialMy wonderful colleagues recently dug into #agentharness and autonomous coding on the latest episode of the Agent Factory. Check it out!
@bishopzareh.bsky.social3/3 Where a file-first Markdown harness actually wins: 11 lifecycle stages, an approval gate each, a staleness formula so no project starves, and nothing but Markdown on disk. Clone it, check it: time2magic.com/agent-harness
@barretblake.devFrom the .NET blog...
In case you missed it earlier...
Build Your Own AI Agent Harness in C#, the MafClaw Live Series
https://devblogs.microsoft.com/dotnet/build-your-own-ai-agent-harness-in-csharp-the-maf-claw-live-series/ #dotnet #AI #csharp #Cloud #agentharness #MicrosoftAgentFramework
@barretblake.devFrom the .NET blog...
Build Your Own AI Agent Harness in C#, the MafClaw Live Series
https://devblogs.microsoft.com/dotnet/build-your-own-ai-agent-harness-in-csharp-the-maf-claw-live-series/ #dotnet #AI #csharp #Cloud #agentharness #MicrosoftAgentFramework #microsoftreactor