OpenAI Launches Agents API for Building Reliable Business AI Agents
- ai-news
- openai
- agents-api
- ai-governance
- ai-policy
OpenAI Launches Agents API — Infrastructure for Reliable Business AI Agents
This week (Sept 10), OpenAI released the Agents API in public beta — an API that brings developers the same harness and infrastructure that powers Codex, the company's autonomous coding tool. The goal: let any business build AI agents that work reliably over hours and days, not just single conversations.
Key capabilities include automatic context management that compacts older information as a session approaches its token limit (so agents can keep working across multiple context windows); smart tool search that loads only relevant tools to save tokens; and native support for parallel subagents that split complex tasks into independent pieces. Developers choose their own compute environment — an OpenAI-managed sandbox or partners like Cloudflare, Vercel, Modal, and others.
Early customer results are notable: SafetyKit reported a 60% reduction in cost per case; Hypha, serving financial-services clients, cut failed agent responses by 86% after separating the harness from the sandbox; and Ciridae saw its evaluation score jump from 0.71 to 0.85. There's no extra fee beyond standard token and tool usage.
This is meaningful infrastructure for any business considering automated AI agents in real workflows — including office management and business automation, where long-term reliability and multi-agent coordination are exactly the core challenge. At RealOffice.AI we track developments like this closely, because they illustrate why robust agent infrastructure — not just a smarter model — is what makes office automation practical in the real world.
Warning: Long-Running Agents Silently 'Forget' Governance Rules
Alongside the launch, VentureBeat published (Sept 13) a sobering analysis for anyone running AI agents on long business processes: as a work session stretches over days and thousands of records, the model gradually 'forgets' governance rules set at the start — without crashing, without triggering alerts, and without anyone noticing until an internal audit exposes the gap. The cause: language models are probabilistic engines, not deterministic databases — as context grows, attention dilutes.
The practical recommendation: don't rely on natural-language prompts to enforce compliance boundaries. Separate critical governance rules into a deterministic policy layer that validates every agent output before it acts, and run mid-task integrity checks (day 4, day 10, day 30) rather than only at the start. Especially relevant for small and mid-sized businesses beginning to deploy AI agents in financial or administrative processes.
OpenAI Launches a Dedicated Financial Services Product
On the same day (Sept 10), OpenAI also announced ChatGPT for Financial Services — a tailored version of ChatGPT Work for banking and finance, developed with design partners Morgan Stanley and Evercore. The product combines GPT-6 Astra with built-in financial data from providers like Daloopa, PitchBook, and Crunchbase, letting bankers build financial models and client materials while tracing every figure back to its source. It includes enterprise-grade security and governance controls, including access management and sensitive-data handling.
The AI Slowdown Debate Reaches US Politics
Two days after Anthropic CEO Dario Amodei called on the industry to "pace the frontier" (Sept 12), former President Barack Obama said publicly that Democrats need to make AI a "central agenda" with a "very clear plan" to address its economic impact and safety concerns. President Trump and House Speaker Mike Johnson pushed back, arguing that slowing down risks ceding the AI race to China. Johnson warned that rushed regulation could be a "national security threat."
The clash shows how the AI-pacing debate has moved beyond the tech industry into mainstream US politics — with potential implications for future regulation that would also affect businesses using AI tools.