Agents
5 items across 1 edition. First seen Fri 11 Sep, last seen Fri 11 Sep.
- OpenAI released the Agents API in public beta on 10 September. It exposes the same managed harness that powers Codex: OpenAI provisions the sandbox, manages session state, compacts context and handles recovery, while the developer supplies tools and tasks.
- Inside a session an agent can execute code, edit files, connect to MCP servers, apply skills, produce artifacts and delegate to subagents, with concurrency capped at 4 by the max_concurrent_subagents setting. There is no separate harness fee; billing is standard model, tool and container rates.
- The docs state the beta currently supports US data residency only and is not eligible for Zero Data Retention, even with self-hosted sandboxes — a material constraint for regulated buyers. OpenAI's announcement post at openai.com blocks automated retrieval, so the figures here come from the developer documentation rather than the launch blog.
- "The Missing Boundary: How Autonomous Agents Lose Control" (arXiv 2609.11024, submitted 10 September) tests five agent models across 16 operational domains and 1,800 trajectories. Neither a degraded control boundary nor an executable unsafe opportunity alone produced substantial loss of control; together they produced a 55% loss-of-control rate, and 62% across ten further domains.
- The paper reports that restoring the original control boundary drops the rate to 0% even when the unsafe action remains executable, and that context compaction is not itself the problem: preserving control constraints through compaction yields 0%, while omitting them raises the rate to 87%.
- This is a preprint and has not been peer reviewed. The environment is deterministic and multi-turn rather than a live deployment, so the absolute rates should be read as a controlled measurement, not an incident frequency. The practical claim — that constraints must survive context compaction — is testable by anyone running long-horizon agents.
- "Big Enough to Break Out" (arXiv 2609.10780, submitted 9 September) compares two PentestGPT-based systems: a legacy human-in-the-loop system on open-weight Kimi K2.5, and a newer autonomous system on Claude Opus 4.8. Across three public targets the autonomous system solves all three, including the two the legacy system never finishes.
- The authors flag the legacy result as the more surprising one: even on machines it fails to solve, it completes about half the subtasks while running on ordinary university GPUs with no provider guardrails — a capability floor available to anyone with open weights and campus hardware.
- The paper explicitly declines to attribute the gain, since model, harness, autonomy and memory architecture all changed together. Adding a coverage-memory layer to both systems improved neither, and in reviewable stalled runs the limiting factor looked like planning and commitment rather than lost memory. Three targets is a very small sample.
- Lt. Gen. Paul Stanton, director of the Defense Information Systems Agency, said on 10 September that decades of delayed maintenance have left Defense Department networks increasingly vulnerable in the AI age, and that the number of zero-day vulnerabilities has "multiplied by a factor of ten." On adversary automation he said: "The ways in which an adversary could employ cyber agents is mind-boggling in terms of the complexity."
- Stanton's stated remedy is to stop deferring patching and operating-system upgrades, treat networks as weapon systems, and train cyber operators on them the way combat troops train with weapons, with validated proficiency standards.
- On defensive AI specifically, DISA intends to require that human operators understand agent behaviour before deployment and to use digital twins to forecast the impact of an agent before it is let loose on a live network — a notably more cautious posture than commercial agent rollouts.
- No budget figures, timelines or patch backlog counts were given in the reporting, so the scale of the remediation task is not quantified.
- Announced 10 September: Skild AI's S1 learns new manipulation tasks from a single video demonstration using in-context learning, with no weight updates or task-specific retraining. NVIDIA reports roughly 66% per-step success on multistep tasks against about 9% for comparable systems, and that one video example is worth roughly 380 hands-on training examples — 50 to 100 hours of manual collection.
- S1 executes unfamiliar tasks up to 10 minutes long across dozens of steps, including potting plants, making pancakes, pour-over coffee and kit assembly. In one plant-potting test, the gap from recording the demonstration to autonomous execution on hardware was 11 minutes.
- Skild reports a $100 million annual revenue run rate ten months after launch and more than 60 deployment partnerships across manufacturing, logistics, inspection, security and food preparation. If the demonstration-efficiency claim holds outside curated tasks, the cost of teaching a robot a new job falls by orders of magnitude — which is the labour-substitution variable to watch.
- These are vendor figures published on a supplier's blog, not an independent benchmark. "Per-step" success is not end-to-end task success, and the 66% versus 9% comparison does not name the baseline systems.