MongoDB Pushes Agent Infrastructure as Search Traffic Keeps Sliding

The announcement came in a MongoDB blog post dated 29 September 2026. It was signed by an executive who said the board had asked him to step in as interim CEO after "this week's leadership change." MongoDB also said it was shipping MongoDB 9.0 and a new deployment option called Atlas Infinite, which lets storage and compute "scale independently as demand changes."

Atlas Agent Engine is neutral across models and frameworks and is built on open standards including MCP and A2A, according to the product page. The company says the runtime traces every model call, tool call and memory access. Policies cascade from the org level and cannot be overridden downstream.

That is the pitch. Whether governance tooling is what enterprises actually buy is unproven. The public preview pricing, MongoDB notes, is "subject to change."

Where the agents actually run

Elsewhere in the stack, the same week produced a cluster of open-source and research work aimed at making agent decisions cheaper and more measurable. On 29 September the OpenJev project published a Jev-compatible decision engine that returns typed answers, Choice, Score or Noul, from open-weight models in a single forward pass, with no JSON parsing. Its README reports 290 ms for one question and 328 ms for 27, with output tokens rising from 23 to 594. The project states plainly that it is independent and "not affiliated with or endorsed by TypeSafe."

Routing is the other half of the cost problem. Unblocked described building an adaptive router for open-weight inference that moved traffic between Baseten, Fireworks and CoreWeave based on measured cost and speed. Under round-robin, Fireworks served 51% of GLM 5.2 tasks and Baseten 49%, even though Fireworks' prices were 25% higher. A fixed order then pushed Baseten to 98.5%. CoreWeave's list prices were about 45% below Baseten's, but the team had no performance data for it.

Two economics papers landed on 29 September with harder numbers. A NBER working paper by Matthew Schwartz, Isaiah Andrews and Jesse M. Shapiro ran an LLM workflow over 4,452 published replication packages from five economics journals and flagged discrepancies in 3,460 articles or their appendices. In 496 articles it cut a calculation's computation time by more than a factor of 10 at similar or greater accuracy, and in 923 it produced an extension absent from the original paper. The paper discloses that Schwartz worked as a contractor for Anthropic and that the results are not endorsed by Anthropic.

The macro claim, and its method

A second paper, NBER working paper 35793, sizes AI's effect through one channel: software engineering. Using stock return sensitivity to an AI index and each firm's share of payroll in software engineering, the authors estimate that from November 2022 to December 2025 AI raised the market's expected present value of software engineering productivity by the equivalent of a permanent 32.6% increase. Their baseline effect on GDP is 3.6%, rising to 6.5% if higher software engineering productivity also lifts R&D productivity. By mid-2026, they write, the effect had more than doubled relative to the end of 2025.

That is a forward-looking measure derived from prices, not a direct productivity count, and the authors present it as such.

On the ground, Picnic CTO Daniel Gebler told The Global Move that his company is "hiring because of the edge AI gives us," not despite it. Picnic runs roughly 400 engineers. Gebler said 25 million of the company's 50 million lines of code were written by analysts rather than engineers. Picnic pulled back on AI-driven release velocity, he said, because 10x more experiments beats 10x more releases.

The traffic side of the story is less encouraging for anyone whose business depends on search referrals. Recent headlines collected for context point to Google search traffic declines steepened to 40% year on year for publishers as of 24 September, with small sites reporting drops of up to 60%. Those are headline summaries, not facts this piece can verify independently, but they frame the week's infrastructure news: the tools being built now assume answers arrive without a click.

Manticore Search made that assumption explicit in a 29 September post. Full-text search still works well for SKUs and exact names, the post argues, but it no longer gets a user to an answer alone. Google reported in May 2026 that AI Mode had surpassed one billion monthly users, the post notes, and people are asking longer questions that never fit a keyword box.