The past ninety days of AI, focused on the scaffolding around the models.
Each new release from frontier model providers garners widespread interest and attention. And while that attention is certainly warranted, there’s been a quieter but equally critical advance over the last ninety days that has happened one layer down, in the harness. We’ve been paying attention to these advancements, as they change how the work actually gets done.
The harness is everything wrapped around the model: tool permissions, the sandbox, memory, the orchestration layer that decides whether one agent or twenty do the work, and the logging that lets you reconstruct what happened. The model supplies the reasoning while the harness decides what that reasoning is permitted to do. It is also the part of an AI system engineers can configure for themselves, which makes its progress this summer worth noticing.
1. Better screening turned into more throughput.
The limit on handing real work to an agent has always been confidence, and the industry’s first answer was the approval prompt. The agent proposes an action, a person clicks yes or no, and the work continues. That checkpoint provides a sense of security, but it means an engineer sits and watches a single agent move through a single task.
Anthropic published data in August that measures what the checkpoint delivers. In a study of 1,053 actions, a classifier screening those actions caught 89 percent of the harmful ones, while human reviewers only caught 13.6 percent initially, falling toward 5 percent after roughly fifty prompts. The study found that users approved 97 percent of what they were shown, signaling that a checkpoint constantly firing quickly leads to muscle memory approvals.
The useful part is what better screening unlocks. When the harness reliably catches the action that would hurt you, an engineer can stop adjudicating individual keystrokes and start supervising outcomes. That is the difference between one agent working under a watchful eye and four working in parallel while the engineer reviews what comes back. Screening quality converts directly into throughput, which is what makes the harness improvements of the past quarter so significant.
It’s important to note that none of these updates hand governance to the tooling. The team deploying the system must still decide what autonomy is appropriate, in which environment, against which rules. What improved is the ability to exercise that judgment at the speed the work requires.
2. The unit of work became a fleet.
Agents now spawn agents. Background subagents run while the primary session continues, sessions on the same machine hand work to each other, and forked contexts let a copy of a conversation take a side task and report back. A job that was one request and one response last spring is now a small team with a supervisor.
That is where the throughput shows up, and it is also where governance gets genuinely hard. Instrumentation built to log a prompt and a completion describes very little about a workflow in which twelve agents touched eleven files over forty minutes. Authority granted to a parent agent has to mean something specific and enforceable about what its children may do. The teams that get durable value out of this generation of tooling will be the ones that can answer, after the fact, which agent did what, under whose authority, and against which version of the rules.
Building with AI under real constraints?
Synergy ECP builds and sustains software for federal customers, which means we put these tools to work inside authorization boundaries, audit requirements, and mission timelines. If that is the kind of problem you want to work on, we would like to talk – head over to our Careers page to check out our openings: Careers – Synergy ECP