Humanity Labs gives wealth management firms an AI Workforce that makes growth profitable. In a people-based business, serving more clients means hiring more people, so cost scales with revenue. The AI Workforce breaks that trap.

It works inside the systems your team already runs and delivers done work across the front, middle, and back office, while your people stay focused on the decisions that matter.

Every job it runs writes into one shared memory, building the firm's Organizational General Intelligence (OGI) so it gets smarter as it grows.

The result is the Firm of the Future: a firm that serves more clients and grows without adding cost at the same rate.

Humanity Labs works with some of the world's leading wealth management firms, which together oversee more than $750 billion in client assets, and is backed by leading venture firms including Google Ventures and Vestigo Ventures.

Learn more at HumanityLabs.ai

Organizational intelligence is the plain question of how a company knows what it knows. Every organization answers it twice: once in the documents it files, and once in the habits, corrections and half-finished procedures its people carry through the work itself. Vendors and researchers spent this week shipping software for the second kind.

Tencent Cloud open-sourced a memory system for engineering teams this week, publishing the stable 2.0.0 release on August 3. It lands on the same premise as a paper Simon Yu, Christopher Manning and five colleagues at Northeastern University and Stanford University posted in May and revised through June: an organization's usable knowledge falls out of finishing work, and the software should record it at that moment rather than ask someone to write it up later.

Why it matters: That paper describes Shepherd, a runtime that records every crossing between an agent and its environment as a typed event, so a run becomes something you can branch from. The recovery problem it solves is specific. "Long-running agents accumulate state that no transcript captures," the write-up explains, and "Git versions files, not a live process or a cache." (Yu et al., May 11) Once the run is a record you can return to, a second agent can watch it and step in before a bad write commits. That supervisor arrangement is what lifted pair-coding pass rates on the CooperBench benchmark from 28.8 to 54.7 percent. An August 8 write-up brought the May work back into circulation as Tencent shipped. (MarkTechPost, Aug 8)

By the numbers: Tencent's system starts from a smaller irritation. "If project context was already explained once, a new session should not need it repeated," reads the release note. (MarkTechPost, Aug 7) Conversations save raw, then an asynchronous pipeline refines them through three further layers until what remains is a scenario or a persona rather than a transcript. A separate asset type, Skill, takes a completed task and writes out the procedure that finished it: versions, resource files, the boundaries where it applies, the steps, and the rules for checking the result. Tencent says accuracy on PersonaMem, a test of how well a system recalls a user's stated preferences across sessions, rose from 48 to 76 percent. Tencent measured that itself, and no independent reproduction has been published.

The big picture: Dwarkesh Patel restated the argument underneath both this week, using saxophone students who may only leave notes for the next player. "There's no sequence of text they could write together that would allow the Nth student outside to play proficiently on their first try," he wrote. (Dwarkesh Podcast, Aug 7) His subject is model training, but the claim lands harder on companies, which have spent decades writing notes for the next player and filing them under knowledge management.

Memory that accumulates from work raises a question a document library never had to answer. A filing cabinet holds what someone chose to put in it. A run trace holds whatever the run touched, so the first time two colleagues share a workspace, somebody has to decide who may read what the other one's agent saw.

Yes, but: Cloudflare found this the hard way. It gave every employee an agent workspace in May, and within three months thousands of staff outside engineering were using it daily. Then people began sharing what they had built. Model Context Protocol, or MCP, is the standard that tells an agent which tools and data it may reach for. "An MCP server can tell you which tools an agent is allowed to call. It cannot tell you which resources that agent already looked at, or where that information went next," Shashi Bellamkonda wrote of the gap. "An agent that reads a sensitive customer table and turns it into a chart has not violated any tool permission. The chart is new. But the chart carries the table's contents." (shashi.co, Aug 8)

Between the lines: Tencent's answer is to make the memory system itself the arbiter. Retrieval-augmented generation, the standard method for letting a model look things up in a company's own files, stops at finding the file. "Standard RAG answers what can be found. The Hub also answers who may use it, which version is valid, and which agent receives it," its documentation says, and its README specifies that a private asset stays owner-only, unreadable even by team administrators. (MarkTechPost, Aug 7) New memories and skills default to private, so a teammate seeing your agent's learning is a decision somebody made rather than a consequence of the plumbing.

What's next: That is a high bar for an organization that cannot answer the prior question. The 2026 Thales Data Threat Report put the share of organizations that know where all their data resides at 34 percent, a figure the knowledge management trade picked up in July. (Document360, Jul 28) Controlling what an agent makes out of a file is harder than controlling the file. Two thirds of these organizations have not finished the easier job.

Ira Kudriashova, who runs software engineering for Jira, interviewed 15 Atlassian engineers chosen by their pull-request throughput and by who their peers named. All of them work on brownfield systems: existing codebases with years of accumulated history and wide surface area, as opposed to something built from scratch. The finding she reports is that AI did not make their fundamentals obsolete; it widened the distance between engineers who practice them and engineers who do not. (Work Life by Atlassian, Aug 9)

Between the lines: The six habits she lists are unglamorous, and several of them store knowledge without anyone calling it that. One engineer rebuilt the test mocks so they behaved like the real backend, and the coverage that produced let AI find and fix most errors without help. Another habit is a "review economy" of trusted, fast, domain-routed reviewers who reciprocate. No one filed either of those as documentation, but both hold knowledge, and both exist because somebody did the work and left the result where the next person or the next agent would run into it.

The big picture: The same organization measured the written half. Tickets composed as executable specifications, with acceptance criteria and out-of-scope notes, scored 4.47 out of 5; the one-line-summary-and-a-link kind scored 2.72. Eighty-three percent of the first group cleared the bar for an agent to act on, against 6 percent of the second, a difference the team reports as statistically significant well beyond the usual threshold. (Work Life by Atlassian, Aug 9) Someone writes a ticket like that while scoping the work, which puts it closer to a run trace than to a manual.

What's next: Fin turned the same loop over to software. Its Operator, generally available this week, reads customer conversations after go-live, works out where the AI succeeded and where it failed, and proposes fixes to help content, workflows and settings. Customers made 20,000 such changes in three months, among them 4,300 pieces of help content written or edited and 3,500 workflow updates. Bob Gerrmann, an AI business process specialist at Road, called it "like having a second brain watching the whole system with you." (CX Today, Aug 10) Every change still waits on a human to approve it.

An agent failure does not look like a failure. "A retrieval step can return the wrong document while every HTTP status reads 200," one survey of the tooling market put it. "An agent can loop through fourteen tool calls, burn thousands of tokens, and deliver a confidently wrong answer." (MarkTechPost, Aug 9) Every monitoring system an organization already owns reports that run as healthy.

By the numbers: Teams have responded by instrumenting. Across all 1,340 respondents to LangChain's survey, 89 percent had some form of observability and 62 percent had detailed tracing. Among the 57 percent already running agents in production those figures rise to 94 and 71.5 percent. Evaluation trails in both groups: 29.5 percent of all respondents run none, and 22.8 percent of the production teams. LangChain fielded the survey between November 18 and December 2, 2025 and published it on June 12, so it describes the ground these teams were standing on before this year's agent releases. (LangChain, Jun 12)

The big picture: Health systems show the same shape from a different angle. A UPMC and KLAS Research study of more than two dozen health system leaders, published on August 6, found more than 90 percent had deployed third-party AI and 92 percent evaluate it before deployment. Only 44 percent have a dedicated data platform for testing. (UPMC Enterprises, Aug 6) The rest check their tools through vendor protocols, limited pilots and informal approaches, which is evaluation of a kind, done without a place to run it.

Yes, but: One organization did publish a grade this week, and it read badly for the humans. Anthropic ran an evaluation across 1,053 paid testers in which a permission prompt was swapped mid-session for a clearly dangerous command. Only 13.6 percent refused it; the company reports its automatic mode would have blocked 89 percent, and it becomes the default for new sessions on Claude Code's Pro, Max and Team plans on August 14. Simon Willison accepted the finding and kept his reservation. "I absolutely buy that auto mode is a better solution than asking humans to constantly approve actions," he wrote. "But... I'd like to see more independent confirmation of this." (simonwillison.net, Aug 8) He then described a malicious-package attack he does not think any version of the feature could catch.

Between the lines: Grindr chose which of its own numbers to publish, which is a problem no observability tool addresses. Grindr chief executive George Arison told shareholders that AI raised engineering output 2.5 times between July 2025 and April 2026, work that would otherwise have taken roughly 200 more engineers and $60 million a year. On the earnings call he said the real multiple was 3.5, and that the company published the lower figure because the higher one sounded unreasonable. (Business Insider, Aug 7) The measure was lines of code shipped, which Goldman Sachs chief information officer Marco Argenti has called a poor proxy.

This magazine, and all the magazines in the Humanity Labs AI Magazine series, were created using Hawkeye, an AI system built for people who work in fields too large and too fast-moving to keep on top of alone.

The stories here are based on a thematic cluster analysis of thousands of articles each week. Those articles range from mainstream news to individual company newsletters. They're analyzed using an inductive reasoning process for trend discovery, a data feature selection process, and a novelty filter for determining which developments are most significant in their respective field.

The analysis is performed and the first draft is written with Claude Opus. It's then fact checked using OpenAI. It's read over and reviewed by series editor and Hawkeye founder Marshall Kirkpatrick, a serial entrepreneur, data journalist, and the first writer hired at TechCrunch. Marshall's writing has been cited twelve times in the WSJ, by ABC, NBC, CBS, the BBC, the American Medical Association, the UN, and many more.