Signals

Signals: Week of November 9, 2026

The week in one read

Multi-agent systems can fail quietly. In one travel-booking setup, agents showed as running while completing zero tasks. In five-agent tests, three small misalignments pushed accuracy close to random. The fixed roster of named agents looks transitional. The pattern looks to be moving toward an orchestrator that staffs each task with fresh sub-agents, with a separate model checking the work and repeat review comments turned into lint rules.

Agents earn trust in steps, from read-only bot to owner of a job, and each one carries its own identity. On cost, the new top-tier model makes the effort setting the price. Persistent agents run $20 to $30 per user per month in compute, which squeezes flat plans. For builders, agents now build niche professional tools and are getting closer to well-scoped contractor work, which opens room for small teams.

The loudest story was NVIDIA losing its CEO as chip export bills gained ground. It may matter for supply, but it likely changes no harness this week. Building looks cheap now. The scarce work may be shifting to defining done and checking it.

Core signals

  1. Multi-agent drift breaks outcomes while dashboards stay green
    Score whole runs on outcomes and alert on drift early
    Agent practice
  2. Orchestrators with fresh sub-agents look set to replace fixed rosters
    Stop building named personas and staff each task per goal
    Agent practice
  3. Checking moves to separate models and into CI
    Turn repeat review comments into lint rules
    Agent practice
  4. Agents earn permissions in steps, under their own identity
    Ship read-only first and scope tables per agent identity
    Agent practice
  5. Effort setting decides cost per task on frontier models
    Default to medium and keep max for whole-project handoffs
    Business-case numbers
  6. Agents now build niche professional tools end to end
    Scope one agent-built replacement for a paid SaaS tool
    Build openings
  7. Always-on agents outrun flat subscription prices
    Test price points and usage caps against per-user compute
    Business-case numbers
  8. OpenAI’s native wallet makes agent checkout a lab feature
    Rebuild agent-payment tooling around the lab wallet
    What platforms absorb
  9. How models describe a brand becomes a sellable surface
    Publish markdown pages and log agent traffic by purpose
    Build openings
  10. Scarce work shifts from execution to intent and judgment
    Write acceptance criteria that agents can check
    Role and skills

Agent practice

Multi-agent runs can look healthy while finishing nothing

Semantic drift between agents breaks outcomes that uptime and per-agent evals never see. In a simple travel-booking setup, agents were up 66% of the time and completed 0% of tasks. Public traces from a multi-agent dev environment showed misalignment in every session. The damage is nonlinear. In five-agent tests, one misalignment cut accuracy from 99% to 98%, two cut it to 81%, and three cut it to 16%. The open-source Scope project reads drift from existing OpenTelemetry traces. Handoffs need alignment at three points: before work (does 100,000 users mean total or concurrent?), during work as scope shifts, and after work against the agreed spec. That last check is still mostly done by a human.
Takeaway
  • Score the whole multi-agent run on outcomes, and drop uptime as a success metric.
  • Make agents confirm the definitions of numbers, dates and goals at each handoff.
  • Alert on drift early. Accuracy holds, then falls off a cliff.

An orchestrator staffs each task with fresh sub-agents

This updates the earlier signal on merging agents into a few owned team agents. A fixed roster of named agents with set roles now looks transitional. Anthropic’s multi-agent research system beat a single agent by about 90% on internal research evals. It sends one agent to a fact check and ten or more to a hard problem, and each gets a narrow task and a clean context. Splitting agents by job function (planner, coder, tester, reviewer) tends to backfire. In one experiment, role-split agents spent more effort coordinating than working. At Uber, most agent sessions are now started by other agents, and a library of 3,600 skills runs over 30,000 times a day.
Takeaway
  • Build the harness around a goal, an orchestrator and short-lived workers over shared company context.
  • Let business events start work (failed CI, renewals, arriving invoices), and keep the skills that handle them in one shared library tracked by revert rate.

Earlier: Merge the agent pool into a few owned team agents; At one commerce platform, engineers operate many agents, and judgment likely stays human

A different model checks the work, and repeat review moves into CI

This updates the earlier signal on code review as the bottleneck. Shopify’s security system treats each finding as a claim. If the agent cannot write a test that exploits the bug, the finding is rejected or downgraded, and a different model does the check. At RunLayer, a fixing agent attaches screenshots or a recording, an adversarial agent reviews, and a QA agent attacks the change after it ships. Heavy review skills can backfire. A /ship skill that checks every changed line burned tens of millions of tokens, and its checklist is being cut from 50 bug types toward 10. At one Electron desktop app, turning a recurring review point into a dependency-graph lint rule in CI removed that class of review tokens.
Takeaway
  • Define proof of done for each task, such as a failing-then-passing test or a recording, and assign a separate checker model.
  • When a review comment repeats, make it a lint rule or CI check.
  • Track tokens per review skill and prune checks that catch no real bugs.

Earlier: Code review becomes the constraint on AI coding

Promote agents like new hires, each under its own identity

This updates the earlier signal on agents acting beyond intended permissions. Trust grows in three steps: read-only access, end-to-end action on a task, then ownership of a job. Most companies are still on step one, for example a read-only data bot in Slack that answers quick analysis requests. Raw MCP database access has no access control, so each agent needs its own identity with an on-behalf-of user, scoped by role and table. Review can scale with risk. RunLayer auto-merges small low-risk changes, and Uber traces every agent-to-agent call to a responsible person. In one early study, managers checked planted errors less carefully when the work was labeled as coming from an AI teammate, and scrutiny dropped about 16% at companies that already treated AI as employees.
Takeaway
  • Ship every agent as a read-only bot first, and write down what it takes to promote it.
  • Keep production writes off until you have audit trails and approval rules.
  • Skip human-style names and personas that invite lighter review.

Earlier: Agents act on permissions their owners never meant to grant

Build openings

Niche professional tools are now an agent build

A top-tier model built a prototype operating system from an empty folder, with one coordinating model driving over 1,000 subagents for 8 days. It boots on a laptop and has a kernel, drivers, file storage, windows, a login screen and an installer. A working video editor run by an agent came out of a separate build, and the team behind it is no longer sure that buying SaaS beats building. Smaller tools are exposed too. Prospect and competitor monitoring that once needed a roughly $1,000-a-month license can be built as a custom agent in minutes. Seat-priced products such as Atlassian, Figma, Monday and HubSpot may face repricing as agents become the main user.
Takeaway
  • Pick one niche SaaS tool your team pays for and scope an agent-built replacement as a test.
  • If you sell per seat, model a case where agents do the work and seat counts fall.

Well-scoped contractor work is now within agent range

The new top-tier model scored 41.3% on the Remote Labor Index, against 21% for the previous Opus. A year ago the best score was 2.5%. The benchmark’s tasks are better defined than typical white-collar jobs. In a bug hunt with 105 planted bugs across two repos, the model found 58, where earlier models mostly found 40 to 45, and it used less than half the turns of the best of them. Service buyers are already moving. Agency work is moving in-house at some marketing departments, and clients push back on price when they know the work is mostly AI-generated. Accounting and legal may be next.
Takeaway
  • List the audits, bug hunts and small builds in your market that an agent can now finish end to end, and price against incumbent service firms.
  • If you sell services, shift revenue toward enablement, and settle AI-use and IP clauses before work starts.

How models describe a brand is turning into a product

A publisher moved to an allow-list. About 70 self-identified agents get in, and roughly 3,000 others a month are sent to a paywall, where none pay. Declared purpose (training, retrieval, indexing) shows which pages each agent reads. That led to an ad product selling structured brand facts, each linked to a source URL. Allowed agents get stripped markdown with the metadata first. Separately, a model misread an email API’s pricing from a JavaScript widget, and a markdown pricing page fixed the answer within days. Measurement is still thin. Audits score visibility, favorability and accuracy separately, and control-versus-exposed tests show a lift on only two months of data.
Takeaway
  • Ask several models to compare your pricing with rivals’, and publish markdown versions of key pages.
  • Log agent traffic by declared purpose in server logs, since web analytics miss it.
  • Treat AEO tracker numbers as directional before committing real spend.

What platforms absorb

OpenAI ships its own agent wallet at zero fee

This updates the earlier signal on merchants splitting over shopping agents. ChatGPT Wallet lets US Pro users authorize agent purchases through a tokenized card on Visa, Mastercard and American Express, with biometric approval on every transaction. Real card numbers never reach merchants or processors. OpenAI built it in-house and did not integrate any of the startups working on agent payments. It takes no transaction fee at launch. Meta’s assistant is reported to plan a percentage of routed transactions plus a $20 subscription tier. Airlines, eBay and Amazon still block bots in purchase flows.
Takeaway
  • If you sell agent-payment or agent-checkout tooling, assume the lab wallet becomes the default and build around it.
  • Use tokenized credentials and per-transaction approval as the baseline for any agent that spends money.

Earlier: Amazon blocks shopping agents while Shopify opens its rails

Assistants may become the OS that products plug into

Most B2B software likely gets its value from the workflows, data and relationships behind the screen. That makes chat assistants look like a new OS or browser for products to plug into. Products whose interface holds nearly all the value look the most exposed. Two tests help. First, if a pitch still works after you add the word yet to what the chatbot cannot do, it is fragile. Second, ask whether the lab as a company could build it, given its regulatory, architectural and incentive limits. For one email API, an MCP server became a real distribution channel. Its remote MCP server roughly doubled its calls every month, to about 1.2 million by the end of August. Getting there meant building OAuth from scratch and growing from about 20 tools to 80+.
Takeaway
  • Estimate how much of your product’s value survives if the interface moves into an assistant.
  • Treat an MCP server as a product with its own OAuth, observability and marketplace listings, and track its calls like a growth channel.

Business-case numbers

The effort setting decides cost per task on the new top-tier model

This updates the earlier cost-per-task signal. The new top-tier model costs $8 input and $40 output per million tokens after a 20% per-token cut. On one benchmark suite, max effort costs $26.54 per task, Extra High $14.83 and High $9.08. The vendor recommends medium as the default. The model’s edge shows on whole projects handed off end to end. On everyday tasks it often matches the cheaper prior model. In a separate test of a top-tier model against a cheaper one on a simple task, the top model took about three times as long and returned about three times as much text.
Takeaway
  • Model cost per task by effort level. Default to medium or high, and escalate only where a failed run costs more than the extra spend.
  • Send well-defined work to the cheaper model, and keep the frontier tier for open-ended project handoffs.
  • Run real tasks side by side on both tiers, comparing time, length and fit.

Earlier: List prices keep falling, and the effort setting decides cost per task

Persistent agents cost $20 to $30 per user per month in compute

This turns an earlier radar item into a number. Always-on agent products run roughly $20 to $30 per user per month in compute. App-builder tools on non-persistent machines run $4 to $5 of VM cost. The gap comes from agents running all day and from users handing them more work over time. GPU prices are rising, and most players are compute constrained. A flat $20 plan for a day-long agent looks tight against these costs. Compute will likely get cheaper eventually, but whether that takes two years or five is open.
Takeaway
  • Test your price point and usage caps against $20 to $30 a month of compute per user.
  • Put metered tiers or caps on heavy always-on users before they set your margin.

Earlier: Persistent agent VMs may cost $20 or more per user per month, against flat $30-50 plans.

Check a simple model or a query before paying for an LLM

For classifiers, recommenders and credit-default models, classic ML often matches or beats LLMs at far lower cost, and its output is deterministic. In a recommender ranker at a workflow-automation company, gradient-boosted trees beat transformer and LLM rankers on learn-to-rank. LLM rankers also cannot guarantee that a recommended item exists in the catalog, even with fine-tuning and RAG. The same logic applies to data questions. Revenue by country is a database query, and running a large model over 8 million Slack messages and emails to get it is the wrong tool. Inference earns its cost on unstructured data that used to be too expensive to process.
Takeaway
  • Before approving an LLM for ranking or classification, require a simple-model baseline and compare cost per request against the metric uplift.
  • Send each task to a query first and to a model second.

Role and skills

Intent and judgment look like the new bottleneck

This updates the earlier signal that review and scope choice may replace coding as the bottleneck. Pull requests grew from about 2,000 to over 4,000 a month on one team, and its agents merge their own. One feature was built in a single public Slack thread of about 200 messages between two colleagues (one a PM) and a bot. Review was done on videos and screenshots. Agents tend to over-build: an agent-provisioned Postgres instance, where serverless DynamoDB would do, added about $300 a month. Teams that hand understanding to agents can build up comprehension debt. In tests of the new top-tier model, its review agents fixed errors but missed work that was boring or irrelevant.
Takeaway
  • Spend PM time on roadmap clarity and acceptance criteria that agents can check.
  • Hire and promote for architecture judgment and the habit of rejecting agent output.
  • Have an engineer explain each agent-built system before it ships, and keep a human on the taste check.

Earlier: Review and scope choice may replace coding as the bottleneck

The AI engineer title is drifting toward application work

Job postings scraped over about nine months across the US, UK, Netherlands, Germany and India show PyTorch requirements for AI engineer roles falling from 40% or more early in the year to under 30%. The title is converging on a software engineer who uses AI tools. The counter-case is that without training, fine-tuning and deployment knowledge, the role is plain software engineering. For hiring, a company-specific prototype or analysis shows AI skill better than a resume line. The candidate builds it before the interview with the strongest reasoning model and checks it by hand.
Takeaway
  • Define each AI role by the work it needs (application building or model training), whatever the title says.
  • Ask candidates what they built for your business, and weigh that above claims of AI fluency.

Backdrop

  • NVIDIA’s founder resigned as CEO after an emergency board meeting, ending a 33-year run. The CFO is interim CEO and no long-term successor has been named. The stock fell about 15% in a week after a November 6 report alleged that a relationship with a Singapore-linked executive helped route an estimated $7.5 billion of chips to Chinese companies since 2023. The incoming House Intelligence chair has questioned the spy claim.
  • OpenAI is backing bipartisan chip export bills aimed at the year-end defense bill, with proposals for chip tracking, diversion alerts and geo-shutdown. Their odds rose after a DOJ criminal probe into Megaspeed, a Singapore compute provider that serves Alibaba, and after Democrats took both chambers in the midterms.
  • Anthropic’s IPO is priced at roughly $1.4 trillion: 825 million shares at $70 to $78, below the $2-3 trillion floated earlier. Its run rate was $81 billion at the end of October, and Q3 revenue was $17 billion against $10 billion at OpenAI.

Where people disagree

Is serving inference a good-margin business?

Inference behaves like a high-margin commodity market, and an open-model lab that owns both the model and a serving stack faster than any rival server for that model keeps healthy margins.
Third-party inference is a roughly $5B market where one platform reached a $16B valuation on over $800M of ARR with an unconfirmed gross margin near 6%, and compute may commoditize within three to five years.

Will inference get more expensive for buyers?

Data center backlash puts an estimated 25 to 35% of next year’s US buildout at risk and could raise license prices for capacity planned for 2028 and 2029, while GPU prices are already rising.
On-demand serving already competes on price, and the two leading labs are close enough in quality that switching is easy.

Is forward-deployed engineering a lasting role?

It looks transitional, since prompts now produce prototypes in seconds and trained business people may build their own tools within a year or two.
Enterprise value is held back by ambiguous problems that someone must define with the business owner, and delivery is moving to system integrators because in-house engineering cannot scale.

On my radar

  • An outside evaluation estimates AI roughly doubled one frontier lab’s R&D pace since October 2025, within a range of 1.1x to 2.9x. Much of the evidence is redacted. If the estimate holds, capability steps may come faster, and forecasters have already pulled their timelines earlier.
  • A $1B Series A for a new lab rests on its claim that one researcher with agents pre-trained a coding model from scratch in 3.5 months for under $10 million of compute. The claim is unverified. If narrow pre-training stays this cheap, small teams could own specialised models.
  • Fully automated loops find the exact error point in agent traces, cluster the failures, and build simulated environments to train on them. This updates the earlier simulated-gym radar. Full traces logged today become the training data for these loops later. Earlier: Simulated data-center gyms where chaos agents break systems and remediation agents learn from the result.
  • Prosumer inference keys come with default spend limits around $100K, and some stolen-key bills reached $1M before anyone noticed. This updates the earlier key-theft radar. Incumbent security vendors are buying agent-security startups at around $400M. Earlier: Stored user API keys from a coding harness were taken by an agent swarm and burned on token generation, hitting 1.4% of console users with bills in the hundreds or thousands of dollars.
  • US new company formations hit about 5.6 million in 2025, and over 70% are expected to stay one-person. Solo operators are a growing customer segment.