Jul 24, 2026 · 2 min listen · Last updated July 24, 2026
From storyflo. This is your daily audio brief. Hey, it's Theo. July 24th. Five things in tech that mattered this morning — let's start with the one that surprised me most. Let's get into it. First, from AI News. OpenAI Presence sells enterprise AI agents with engineers attached.
Listen · storyflo · A.I.
Daily A.I. Brief · July 24th
0:00-2:16
Pick your daily storyteller
Subscribe to match with Theo, Jessica, Chloe, Mason, Brock — your voice, every brief.
Audio pre-rendered by Storyflo · cached + delivered from the edge
OpenAI Presence sells enterprise AI agents with engineers attached
OpenAI’s new Presence isn’t a product you click‑order; it’s a managed engagement where their Forward‑Deployed Engineers (and a few partner integrators) roll out a custom AI agent for a single business task—think handling a billing dispute or an IT ticket. The agent only gets the data and system access it needs, and the client writes the rules that dictate when it can act, when it must pause for sign‑off, and when a human steps in. After launch, Codex watches the sessions, flags oddities, and suggests tweaks that the client’s team can test before rolling out.
OpenAI lays out a six‑stage workflow: scope the outcome, run security and legal reviews, simulate, accept‑test, staged rollout, then iterate post‑launch. Guardrails and escalation paths are baked in, and every interaction is logged for audit. The model itself isn’t fixed; it’s configured per workflow and can evolve, which means contracts need to spell out what version the agent is “held to.”
Because the service relies on engineers embedded in the customer’s environment, availability is limited and pricing stays private. Early pilots with BBVA, SoftBank and IAG are still in design phases, not full‑scale deployments, so the real test will be how the managed approach scales once the consulting capacity runs thin.
Language Model Hallucination Evaluation with GraphEval
Turning the key principles and methodological stages of GraphEval into a simulated practical scenario to better understand its usefulness and key implications in understanding and combating LLM hallucinations.
Kimi K3 trails frontier US models by a wide margin on cyber exploits, and distillation may explain why
I just saw the latest test on Moonshot AI’s Kimi K3 and thought you’d want the details. The British AI Security Institute teamed up with the U.S. Center for AI Standards and Innovation to throw a suite of offensive cyber tasks at Kimi K3. On ExploitBench, which measures how well a model can generate or recognize exploit code, Kimi K3 landed at about 32 percent, while the top U.S. models were cruising around 76 percent. Even more striking, its safety layers didn’t stop the model from actually building or simulating attacks.
What’s odd is that Kimi K3 still scores solidly on the usual language benchmarks, so the dip is specific to the cyber‑security arena. The researchers suspect the cause might be “distillation” – essentially a streamlined copy of Anthropic’s models that lost some of the nuanced defenses built into the original. If the distillation stripped away those safety nets, it could explain why the model looks strong on general tasks but falters when you push it into exploit territory.
Claude's voice mode now runs on Anthropic's most capable models across all platforms
Voice conversations now run on the more powerful Opus and Sonnet models with access to Gmail, Google Calendar, and Slack. Claude is currently the only AI assistant that can compose and send emails directly by voice, giving it an edge over OpenAI and Google, whose voice output still sounds more natural.
The article Claude's voice mode now runs on Anthropic's most capable models across all platforms appeared first on The Decoder.
Context Windows Forget What Matters — I Built a Usage-Reinforced Decay Engine for AI Agent Memory
I’ve been thinking about how LLMs just keep the newest bits in their context window, even when older facts are the ones you actually need. The author took the classic Ebbinghaus forgetting curve and turned it into a usage‑reinforced decay engine, so memories fade unless they get pulled up repeatedly.
Instead of a flat sliding window, the system tags each piece of information with a decay timer that resets every time the agent references it. When the timer runs out, the data drops out, making room for newer, but still relevant, context. The trick is that the decay rate adapts based on how often the memory is accessed, so the most useful facts linger longer.
In practice, they hooked this into a simple AI agent and ran a few benchmark tasks. The agent kept the right details longer, which trimmed down the amount of prompting needed and cut down on hallucinations. It’s a neat way to let the model “forget” intelligently rather than just by recency.
Loop Engineering for RAG Generation: An LLM Cascade from a Cheap Local Model Up to a Hosted Flagship
Enterprise Document Intelligence [Vol.1 #8quater] - Two angles on the cascade, cost and a validation loop, backed by a real sweep of twenty local models against a hosted flagship
The post Loop Engineering for RAG Generation: An LLM Cascade from a Cheap Local Model Up to a Hosted Flagship appeared first on Towards Data Science.