mcp 0.5.0 - does agent see tool call history from previous turns
tested this with a postgres tool that queries user data across multiple turns. question is: can the agent see the full history of previous tool calls (params + responses) or does it only see the current conversation? iirc the conversation context should include tool history but want to confirm before we rely on this for multi-step workflows. could be wrong tho
does it keep the full history or just last N turns? worried about context limits if i'm doing long debugging sessions
yeah it keeps full history. we tested this with a github tool that references previous PR numbers and it works. just be careful because context gets huge after long sessions - we hit 45k tokens after 30 tool calls
45k tokens after long sessions is rough. do you just clear history periodically or is there a way to compress previous turns? what's the actual context limit before the agent starts losing track of early tool calls?
yeah it keeps full history which is great for our use case (internal debugging tool that needs to reference previous tool calls). the context does get pretty long after 20+ turns though so we added manual truncation logic to keep the last 50 tool calls. does anyone know if there's a built-in way to configure history retention or do you just have to roll your own?
wait so fable 5 is like the premium one but it got nerfed so now everyone uses sonnet 5??
Roughly linear in our tests but with enough noise that you need to sweep a 4-6 layer window to find it. On Llama 3.1 14b we found the equivalent composition spike at layer 13 head 4, which tracks with the linear interpolation between 8b (layer 11) and 70b (layer 18). The coordinates aren't perfectly predictable though - architecture details like attention head count and MLP width ratio affect where the pattern lands.
yeah it does. agent sees full tool call history from previous turns as context. we tested this with a stripe tool that needs previous charge_id from earlier calls and it works fine