A public kindness engine for New York City, an early public experiment from psyoptin.com. The engine finds and prepares; people bring judgment and action.
In an experiment Anthropic published on 2026-04-24, 69 of its own employees were each given about $100 and an agent, to buy and sell from one another through the agents. They made 186 deals and moved just over $4,000. The finding that travels: people whose agents ran the stronger model got better prices, and the people whose agents ran the weaker model did not notice they were at a disadvantage when they rated the deals afterward. Anthropic called it Project Deal. It is small, but it used real money and real people.
This dispatch is the practical one: where agents are used in production and with what evidence, eight products worth knowing, and what to be careful of.
This is the strongest signal and the most argued-over number. Cognizant, Cognition and Odyssey Logistics announced on 2026-09-23 that autonomous engineering agents in production produced a "37 percent net cost saving" against "the conventional approach"; the release, from companies that sell the service, gives no method or baseline. METR's 2025 randomized trial found experienced open-source developers were 19 percent slower with these tools than without. Its follow-up, published 2026-02-24, produced intervals that span zero, and METR called its data "only very weak evidence" because of selection effects, including developers declining to take part without the tools. The largest number here is a company claim with no stated method; whether agents save money is something each company has to measure itself.
Salesforce reports Agentforce annual recurring revenue "exceeded $1.5 billion, up over 240%" in the quarter it announced 2026-08-26, noting that, effective that quarter, the figure includes Slackbot and Headless 360, which makes it a different measure from earlier quarters' figures. Gartner forecast on 2025-06-25 that over 40 percent of agentic projects would be canceled by the end of 2027. Cognizant's 2026-09-28 announcement of agents for pended health-insurance claims in its TriZetto system is a pattern worth noticing: early-adopter health plans are piloting it, and Cognizant says performance results will come in a future update.
Vendor figures for customer-support and sales-outreach agents could not be traced to an independent source for this dispatch, so none are repeated here. On the phone, Instinct, the personal agent that raised $1 billion on 2026-09-28, offers a concierge that phones businesses without online booking. 404 Media reported that Meta's Muse was tested internally with a "human agent layer," meaning a trained human placed some calls, and that one employee tester was told only afterward that the caller was human; a Meta spokesperson said it will roll the feature out only when it is ready and with the proper disclosures. When a product says its agent makes calls, ask who is actually on the line.
Code for America and Anthropic announced a SNAP Policy Navigator pilot for benefits caseworkers on 2026-05-08; the announcement reports no results. The MyCity chatbot, which The Markup and THE CITY caught in 2024 telling businesses they could take workers' tips, was announced by Mayor Mamdani on 2026-01-30 as ending and taken down 2026-02-04; The Markup put its cost at about $600,000 to build and $500,000 a year to run. Route Fifty's July report, which covers tools used during 2025, describes narrower ones: the 311 app tries to match an uploaded photo to relevant service requests, and the youth-services department's virtual assistant handles about 2,100 requests a day. A state comptroller's audit, reported by Route Fifty in September, says the city still lacks a complete inventory of these tools.
| Name | What it is | Evidence status |
|---|---|---|
| OpenClaw | Open-source always-on personal agent with memory files | 247,000 GitHub stars by 2026-03-02 (Wikipedia) |
| Meta Muse and Sentinel | Consumer agent on its own cloud computer, with one permission broker | Company description; no independent audit found |
| Instinct | Text-and-phone-first personal agent, invite-only | Funding per TechCrunch; no user numbers disclosed |
| Cloudflare Web Bot Auth | Agents sign each web request; sites set per-agent policy | Launched 2025-08-28; built on an IETF draft |
| MCP, goose, AGENTS.md | Tool protocol, agent runner, and a README for agents | Under neutral Linux Foundation governance |
| Project Deal and Project Vend | Anthropic's experiments with real money and staff | Published methods; small samples |
| CivicMesh | Hackathon navigator for housing, food, health and legal aid | Its README: hackathon winner, 40 resources, live demo |
| civic-tech-agent-bridges | Command-line bridges to Pol.is, Decidim, CONSUL | Brand new; 0 stars, 7 commits when seen |
OpenClaw. Austrian developer Peter Steinberger's assistant launched in November 2025, was renamed Moltbot in January 2026 after trademark complaints from Anthropic, then OpenClaw days later, and had 247,000 GitHub stars by 2026-03-02, per Wikipedia.
Muse's Sentinel. One permission broker sits between the agent and every connector and network call; the agent holds surrogate tokens, and real credentials are swapped in at the network boundary. That is Meta's own description, as relayed by MarkTechPost; this desk found no independent audit, and a researcher published a proof of concept against the Mac app (below).
Instinct. Its distinctive idea is a trusted-person network so agents can coordinate between friends. It also shows that investors are willing to value a product at $10 billion with no published user count.
Web Bot Auth. An answer to "who is this bot." Agents sign their requests so a site can recognize them and set a policy for each.
MCP, goose and AGENTS.md. The protocol most agents use to reach tools, Block's open-source agent runner, and OpenAI's AGENTS.md file are now under a Linux Foundation body announced in December 2025, with the major labs and cloud companies as members.
Project Deal and Project Vend. Vend's second phase reports a Claude-run shop moved from losses to profit after it got a customer-record system, other tools and a "CEO" agent; the authors caution it may have happened in spite of the CEO agent. Both projects are small, and both publish how they were done.
CivicMesh and civic-tech-agent-bridges. Two small public code projects. CivicMesh, which its README says took first place in an agentic track at JacHacks Spring 2026, turns one message into a ranked plan across housing, food, healthcare and legal aid from 40 resources. civic-tech-agent-bridges is an Apache-licensed set of command-line bridges so an agent can operate participation platforms like Pol.is, Decidim and CONSUL. Neither has been independently evaluated; both are built in public, so anyone can check what they do.
The habits that fall out of all this: give an agent surrogate credentials, keep secrets away from anything that reads untrusted text, keep a log the agent cannot edit, and put a person before any step that cannot be undone.
The products that deserve attention this month are not the ones with the largest claims. They are the ones that publish how they work, sign their requests, hold credentials at arm's length and let anyone read the code. That is a reason to measure your own use before trusting anyone else's number.
What to check for yourself: open Project Deal and read the method, then ask whether any vendor case study you have been shown does the same. Look up whether an agent you use signs its requests or can be identified by the sites it visits. And if it installs skills or plugins, find out who audits them, because in at least one documented case the answer was a security firm, after the malicious ones had already been published.