SUBJECT Signal Brief: agents get wallets, self-improvement, and a flat-rate model plan
Composio Instant
Composio Instant — launched October 20 — gives an AI agent its own wallet so it can pay for the tools it needs itself — web search, business-contact lookup, voiceovers, audio transcription — with no subscriptions or signups, free to try.
WHY IT MATTERS Agents that can transact for themselves cross a real product line — agent-held spend with per-call pricing and no human checkout is the missing piece between a demo agent and an autonomous workflow, and it forces PMs to spec budgets, limits, and audit trails for machine customers.
composio.dev →Harness²
Harness² — launched October 20 — is Google Research's open-source recursive agent harnessing framework that lets an LLM agent evolve its own execution harness — prompts, tools, and memory — without ground-truth verifiers, reporting up to +11.5 points on professional benchmarks like WorkBuddy-Bench.
WHY IT MATTERS Self-improving agents move prompt-and-tool tuning from a recurring PM chore into the product itself — a plug-and-play layer that compounds an agent's harness over time changes what teams hand-optimize versus what they let the system learn.
github.com →Exa 3.0
Exa 3.0 — launched October 20 — is the biggest search-quality update to Exa's search API for AI agents: a larger embedding model, billions of new documents including 1.27 billion ecommerce listings and 211 million-plus local businesses, and Dynamic Highlights that use 95% fewer tokens than full page text — from $4 per 1,000 requests on the new Lite tier.
WHY IT MATTERS Retrieval cost is becoming the line item that decides which agent features ship — a quality jump at a $4-per-1,000-requests tier with 95% token savings on highlights re-prices what always-on search inside a product can afford to do.
exa.ai →RunInfra
RunInfra — launched October 20 — sells flat-monthly coding plans that put fast open models — GLM 5.3, DeepSeek V4.1 Flash, Qwen3.8 27B — inside Claude Code, Codex, OpenCode, Cline, Droid, or Grok Build, with one-prompt setup, session-aware routing that hits cache up to 99.5% of the time, zero data retention, and SOC 2 Type II, starting at $10.
WHY IT MATTERS Flat-rate open-model plans turn agent inference from a metered anxiety into a predictable budget line — and a bring-your-own-harness model router is a direct signal that the coding-agent platforms' grip on model choice is loosening.
runinfra.ai →ToDo.is
ToDo.is — launched October 20 — is an always-on AI agent you hand to-dos to: it researches with sources, makes files and images, uses websites, and reports back on web, Android, WhatsApp, or Telegram — running around the clock in its own sandbox with scheduled briefings, and it asks before it pays, sends, or deletes anything, with a $1 trial.
WHY IT MATTERS Consumer-grade agents are settling on the same contract enterprises want — persistent memory, scheduled proactive work, sandboxed execution, and human approval before anything irreversible — which is fast becoming the baseline checklist PMs will be measured against.
todo.is →