Releases·Tool Release
Ex-Sierra agent builder's Charter library starts every tool as a word-for-word API copy, then cuts it down
A developer with two years of production agent experience (Surfbuddy, now archived) released Charter, an Apache-2.0 open-source library that declares tools as complete word-for-word copies of API definitions, then trims them down. The approach surfaces issues like oversized tool outputs and lets rules be attached as per-field glosses. The library ships with 15 pre-declared APIs including Gmail, Slack, Stripe, Linear, and Notion.
8 engagement1 sourceThu, Oct 1, 2026, 03:13 PM
AI Overview
- Surfbuddy agent in real users' Gmail, Calendar, and Sheets found Gmail's threads tool returned 314 KB per call; parsing cut it to 2.4 KB before the agent sees it
- Charter author advises declaring tools 100% first, word-for-word from API docs (delegable to a coding agent), then trimming, so cuts and undos are one line
- Field-level glosses beat system-prompt rules, which break; a gloss fixed a 3B-active Nemotron refunding $0.15 on Stripe instead of $15
- Charter is Apache-2.0 open source with 15 APIs declared, including Gmail, Slack, Stripe, Linear, and Notion
- Quotes Sierra's agent development lifecycle write-up on prompt changes breaking prior fixes, 'a game of prompt engineering whack-a-mole'
Related
Social·Community
As agents gain more tools, direct tool calls break down; a sandbox where the model writes code to call tools fixes it
A discussion post describes how agents degrade as the tool list grows: schemas fill context and the model mis-calls or clumsily orders tools. The fix that holds up is giving the model a sandboxed code environment where it calls tools as functions, chaining and filtering them in code. This cuts token use per run and reduces wrong calls, at the cost of securing a sandbox running model-written code.
A discussion post describes how agents degrade as the tool list grows: schemas fill context and the model mis-calls or clumsily orders tools. The fix that holds up is giving the model a sandboxed code environment where it calls tools as functions, chaining and filtering them in code. This cuts token use per run and reduces wrong calls, at the cost of securing a sandbox running model-written code.
49 engagement1 sourceredditThu, Oct 1, 2026, 02:23 PM
Social·Community
Practitioners report tool bloat degrades agent performance
Developers report that giving agents more tools degrades performance, with tool lists of 15-20+ entries leading to wrong tool selection and wasted turns. Some note that Genie Code caps MCP at 20 tools, suggesting a measured limit. The community recommends pruning tools or using sub-agents to reduce context load.
Developers report that giving agents more tools degrades performance, with tool lists of 15-20+ entries leading to wrong tool selection and wasted turns. Some note that Genie Code caps MCP at 20 tools, suggesting a measured limit. The community recommends pruning tools or using sub-agents to reduce context load.
340120 engagement7 sourcesreddit, rssWed, Aug 12, 2026, 04:16 AM
Social·Community
I spent days fixing the prompt when the real problem was the tool schema
A developer spent days adjusting system prompts to fix an agent repeatedly choosing the wrong tool, before discovering the tool descriptions were the actual problem. The descriptions, copied from internal API docs, explained what functions did but not when to use them. After rewriting the tool schemas with clearer usage guidance and explicit parameter constraints, tool selection accuracy improved immediately with minimal prompt changes.
A developer spent days adjusting system prompts to fix an agent repeatedly choosing the wrong tool, before discovering the tool descriptions were the actual problem. The descriptions, copied from internal API docs, explained what functions did but not when to use them. After rewriting the tool schemas with clearer usage guidance and explicit parameter constraints, tool selection accuracy improved immediately with minimal prompt changes.
14 engagement1 sourceredditMon, Sep 14, 2026, 01:20 PM
Social·Community
Developers argue 1:1 API wrapper tools cause agent failures
A developer argues that many agent tools are just 1:1 API wrappers, pushing control flow into the model and causing failures. They advocate for higher-level tools like upsert_contact instead of separate search/create/update calls.
A developer argues that many agent tools are just 1:1 API wrappers, pushing control flow into the model and causing failures. They advocate for higher-level tools like upsert_contact instead of separate search/create/update calls.
20 engagement1 sourceredditWed, Aug 5, 2026, 11:33 PM
Products·launched
Worlds: local replicas of Stripe and Zendesk for testing AI agents
Worlds provides local replica environments of services like Stripe and Zendesk so developers can test AI agents against realistic third-party APIs without touching production accounts or spending real money. The agents interact with these sandboxed replicas to evaluate how they handle common integration scenarios.
Worlds provides local replica environments of services like Stripe and Zendesk so developers can test AI agents against realistic third-party APIs without touching production accounts or spending real money. The agents interact with these sandboxed replicas to evaluate how they handle common integration scenarios.
2 engagement1 sourcehackernewsTue, Sep 29, 2026, 05:52 PM