DROPGoogle launches Gemini 3.8 Flash TTS and Flash-Lite TTS///DROPGemini Omni brings 1080p generation to Google Vids///BREAKClaude Opus 5.5 is here: 1M context, $4 input / $20 output///BREAKOpenAI adds GPT-6 Sol and Luna to the GPT-6 family///PAPEROpus 5.5's system card documents gains—and evaluation limits///BREAKGrok 4.7 launches at $2 input / $6 output per million tokens///DROPGrok Voice Transcribe 2.0 targets difficult real-world audio///TRENDGoogle steers new Gemini API projects to 3.8 Flash or 3.5 Flash-Lite///DROPGoogle launches Gemini 3.8 Flash TTS and Flash-Lite TTS///DROPGemini Omni brings 1080p generation to Google Vids///BREAKClaude Opus 5.5 is here: 1M context, $4 input / $20 output///BREAKOpenAI adds GPT-6 Sol and Luna to the GPT-6 family///PAPEROpus 5.5's system card documents gains—and evaluation limits///BREAKGrok 4.7 launches at $2 input / $6 output per million tokens///DROPGrok Voice Transcribe 2.0 targets difficult real-world audio///TRENDGoogle steers new Gemini API projects to 3.8 Flash or 3.5 Flash-Lite///
DARC.DEV

[ S02 — The Signal · web edition · 23 Sept 2026 ]

Agents are getting more capable. Make the boundaries more explicit.

Cursor's longer-running projects, MCP's current authorization rules, approval gates and budget enforcement that survives a restart.

← All web editions

The useful agent question is shifting from 'can it call a tool?' to 'can the system explain, authorize and recover from what happened?' This edition connects the current documentation to four concrete engineering checks.

01 / Longer projects need explicit authority

Cursor's current changelog describes persistent Projects, subscriptions and self-hosted execution. Those capabilities expand useful workflows, but shared context and background triggers need ownership, scope and a stop condition.

One practical moveWrite an execution brief before enabling recurring work.

02 / MCP is not an authorization shortcut

The July 28 specification hardens authorization and distinguishes HTTP flows from STDIO credentials. Tool annotations do not replace access checks, and tokens for another resource must not be passed through as MCP authorization.

One practical moveDraw the trust boundary from user to client, server and downstream service.

03 / A resumed agent can repeat a write

Checkpointing preserves internal state. It cannot automatically prove that an external action was executed once. A crash between sending a request and persisting its result needs idempotency or reconciliation.

One practical moveTest a crash immediately after an external write.

04 / Read the budget deployment requirements

LiteLLM's deployment guide explicitly warns that budget enforcement depends on a connected database on the documented path. A configuration value is not evidence that an over-budget call will stop.

One practical moveCreate a tiny staging cap and prove the next request is rejected.

The goal is not less capable agents. It is smaller uncertainty about their authority, spend and effects. Start with a deny-path test and keep the result beside the happy-path demo.