1.9k
OpenAI·2d agoAgents
OpenAI announces agent operator API
Agent Operator is a hosted runtime for multi-step workflows. It bundles browser, terminal, and document tools, plus a memory layer. Pricing starts at $0.05/action.
News
20 articles
Agent Operator is a hosted runtime for multi-step workflows. It bundles browser, terminal, and document tools, plus a memory layer. Pricing starts at $0.05/action.
The flagship model now handles 1M tokens, adds structured output improvements, and reduces tool-call latency by ~28%. Priced at parity with 4.6. Early benchmarks show meaningful gains on long-horizon agent tasks.
The open source agent runtime launches a remix-friendly skills directory. Skills can be installed into any OpenClaw workspace in a single command and paid authors get a 70/30 split on purchases.
Escrow accounts can now hold buyer funds up to 90 days, release on a programmable condition, and issue automatic refunds on dispute. Pricing: 0.4% plus the standard card fee.
Hermes 1.0 is now GA. Memory packs come with episodic, semantic, and procedural layers, and integrate with OpenClaw in under a minute.
A new wave of marketplaces — led by BusellAI and a handful of European peers — are moving $4M+ in monthly GMV by listing cash-flowing agent businesses with verified revenue.
Operators of general-purpose AI systems above the compute threshold now face annual disclosure obligations. Agent operators below the threshold mostly need to log training data provenance.
A new report from Pitchbook shows AI-native SaaS acquisitions climbed 3.2x YoY. Multiples for seed-stage cash-flowing agent businesses now sit at 18–24x MRR.
A team at DeepMind publishes a web-agent that scores 78% on WebArena — up from last year's 52% leader. Paper breaks down the changes in exploration policy.
Paper shows 3B student models can reach 90% of GPT-4 quality on narrow domains when coached by a larger model during training.
Of 240 companies in the latest batch, 62% describe themselves as agent-first. Average round size sits at $1.5M with median valuation near $18M.
A new long-form guide covers schema design, reliability patterns, and how to audit tool calls in production. Includes a working example repo in TypeScript.
Agent builder studios — firms that ship turnkey agent businesses — are raising 5x last year's volume. BusellAI-affiliated builders placed top-5 in three separate deals.
The new open-weight model outperforms prior 7B leaders on MATH and GSM8K, and nearly matches GPT-4-mini on long-context reasoning.
pgvector + Supabase indexing now handles 1B rows with sub-50ms latency on hybrid search. The upgrade is free for all Pro customers.
New templates support approve/reject interactions natively. Agents can now request human confirmation without drafting a conversational fallback.
LangGraph Studio adds live state inspection, tool-use replay, and a new `trace-diff` view for comparing two agent runs side by side.
New diffing view makes it easy to see what changed between two eval runs — including prompt variants and tool-call order.
Led by Andreessen Horowitz. BuildOps ships a hosted agent runtime aimed at SMBs. It runs 200+ production agents today.
Several jurisdictions are examining KYC practices at AI business marketplaces. Most operators — including BusellAI — already require Stripe-grade verification.