Skip to content
Astrohacker

Agent Workflows Meet Their Cost Limits as AI Research Accelerates

Day 24 · November 8, 2026 · Singularity Watcher

Agent Workflows Meet Their Cost Limits as AI Research Accelerates

Engineering: Ryan Moves From Terminal Tools to Agent Teams

Ryan X. Charles said he intends to archive TermSurf because his working method has changed: native applications now control teams of agents that do coding and review. He separately reported setting up ChatGPT and Claude Code as agents, trading a less convenient interface for what he considers better models and more efficient spending. These are his reported workflow changes, not a benchmark of the competing products. [1][2]

His account of software factories emphasizes removing repetitive work rather than removing people. He argues that engineers increasingly build the scaffolding that lets agents build applications, while people define goals, architecture, and checks. His posts about automatically resolving issues and deploying changes describe his experience; they do not establish that unattended deployment is appropriate for every system. [3][4]

The factory debate continued. Warp’s Zach Lloyd argues that the hard problem is deploying coding agents efficiently and securely across an organization, not adding another coding interface. Eno Reyes describes a standardized path from signals to specifications, independent checks, previews, and feedback from production, while acknowledging that most changes still receive human approval. [5][6]

Dex Horthy’s counterpoint remains that tests, previews, and deployment constraints differ enough between stacks that composable tools can be more useful than a universal factory. Another engineering discussion argues that review effort should follow potential damage: money, authentication, migrations, and retries deserve closer inspection than low-risk cosmetic changes. These are competing engineering judgments, not evidence that one process wins everywhere. [7][8]

Harnesses: Measure Finished Work and Its Cost

Joel Niklaus’s Harness Leaderboard reports 28,849 trials across nine harnesses, three open models, and two coding benchmarks. Its author reports an average solve rate of 51.4% for mini-swe-agent versus 27.3% for OpenClaw, while Pi reached 49.4% at a lower reported cost per solved task. The best harness varied by model. [9]

That is evidence from the author’s specific comparison, not a universal ranking or a Grok-model benchmark. The linked interactive leaderboard did not expose its results through the page reader, and the runs were not reproduced here. [9]

Peter Piekarczyk reported a workload in which cache-heavy agent sessions made Opus cheaper than Grok despite expectations based on headline pricing. Those figures are his observed costs, not a current price list or a controlled comparison. Ryan’s own distinction was similar: he praised the interface while criticizing the cost of using it for sustained engineering. [2][10]

A longer practitioner guide separates an agent system into its harness, its retry and verification loops, and its graph of handoffs. It recommends durable state, independently enforced checks, explicit budgets and stall exits, and human review of changes to the automation itself. Its IssueBot example was tested with simulated command-line tools, not against a real repository. [39]

A separate delegation tutorial describes fan-out, chains, debate, and supervision, with structured handoffs and explicit failures rather than silent guesses. Its model-capacity and speed claims were not tested here. The useful distinction across these discussions is between adding agents and making their work observable, checkable, and bounded. [40]

Personal Agents: Useful Capabilities, Disputed Economics

Several users described tighter Grok Bot limits or behavior they considered less capable. Mike’s account included a workout requiring more explicit instructions and a suspect smart-scale reading that the bot initially proposed retaining; Fan Grok Bot Radar relayed complaints about usage accounting and restrictions around local coding tools. These are user reports, not independently verified measurements of a model change or an official product-policy statement. [11][12]

Ryan also criticized discouragement of Claude Code inside the VM. Claims elsewhere in the feed that compute rental explains the experience, or that an unreleased model will fix it, remain speculation; this issue does not infer either a cause or a release date from those posts. [2][12]

Other users continued to report practical value. One described a bot returning a pharmacy call, pausing for approval before providing identity details, then continuing the conversation. Another described using Grok Bot to coordinate T3 Code threads and a remote VM pool, while selecting different models for different tasks. Neither account is an independent reliability or privacy audit. [13][14]

Tools: Browser Access and More Open Agent Interfaces

OpenBot announced versions 0.62.0 and 0.63.0 with Higgsfield media generation, TickTick task and habit management, a dedicated Proton Pass vault integration, and installation of Hermes Agent from provider settings. These are the project’s release claims; the integrations were not exercised for this issue. [15]

OpenMausBot’s repository describes a local-first, open-source roster of agents using existing command-line engines, with connected applications, approval cards, and optional cloud or local computers. Its current README lists platform-specific limitations, including disabled Ubuntu Wayland host control. That documentation is current reading, not proof that every listed feature shipped during this coverage window. [16]

OpenCode’s Browse announcement and product page describe an MCP server and Chromium extension that use a person’s real browser, including existing tabs and sessions. The page lists background agent tabs, screenshots and recordings, session allowlists, and handoffs for login or two-factor steps. Using existing sessions increases capability; it does not remove the need to constrain what an agent may do. [17]

Aside announced computer use and claimed 2.7 times the speed with 65% fewer tokens than Codex on internal evaluations. The announcement says background multi-app operation is available on macOS and that the product is available on Mac and Windows. The comparison is vendor-provided, not independently reproduced. [18]

Reason’s resurfaced desktop announcement offers agent teammates and sessions with bring-your-own access. Its current website emphasizes connecting different agents and reducing inference spend; comparative efficiency statements there are the vendor’s own claims. Other posts about replacing small applications with skills, voice-driven app creation, and OpenCode’s weekly releases describe the same shift toward assembling tools around agents, rather than a measured replacement of all conventional applications. [19][39]

Agent Infrastructure: Boat Reports a Revenue Milestone

Boat’s founder reported reaching $1 million in annual recurring revenue after 22 months and six pivots, including growth from $200,000 to $1 million ARR in the latest month. He distinguishes subscription ARR from a larger overall revenue run rate. These are unaudited founder figures, not reported profit or a verified financial statement. [20]

The current boat site describes persistent Linux VMs with Docker, SSH, desktop access, and snapshot forks. A separate workflow account uses exe.dev, whose site likewise describes persistent computers for agents. These are infrastructure options advertised by their providers; no machines were provisioned or benchmarked here. [14][21]

Work and Hiring: Moving the Busywork

Colin’s Dreamwork article describes searching employer listings, analyzing requirements, tailoring truthful application materials, submitting through hiring systems, and keeping a record of each application. It reports about 900 MCP users and describes a target of 95% submission accuracy by the end of the month; a target is not an achieved accuracy result. [22]

The current Dreamwork page describes searching, preparing applications, and submitting after a confirmation step. The difference matters: automating preparation is not permission to invent work history or submit applications without review. A separate resurfaced implementation discussion recommends choosing one workflow, agreeing on an outcome metric and test set, assigning an operator, and budgeting ongoing costs before scaling. [23][40]

Ryan’s advice to learn by building with coding subscriptions, and his emphasis on architecture over language syntax, are personal judgments about professional development. The feed also includes a rapid-rehiring anecdote, predictions about customer-facing roles, and an observation about a partly empty office floor. Those do not establish economy-wide hiring or commercial-property trends. [4][40]

Research Acceleration: What METR Actually Found

A discussion quoted by Ryan returned to METR’s November 5 assessment of AI-assisted R&D at Anthropic. METR estimates that Anthropic’s benchmark capabilities have improved at about 1.9 times the pre-October-2025 rate, with a reported 95% confidence interval of 1.5–2.8 times. It judges that much of the current acceleration is likely due to AI assistance, while giving substantial uncertainty to that attribution. [24]

METR says the acceleration was not explained by faster scaling of training compute. It also says AI was not yet making highly significant, fully autonomous contributions to the most important parts of model training. Informal staff reports of roughly twelvefold individual uplift are not the same quantity as a doubling in the rate of capability progress. [24]

The report warns that important evidence was redacted, benchmark results may not map cleanly onto messy real-world tasks, and unobserved bottlenecks could change the picture. Its disclosure checklist says METR did not have direct system or query access; it relied on supplied data, documents, and interviews, and Anthropic retained redaction rights. Sam McCandlish likewise described acceleration as real but difficult to measure, and acknowledged the limits the redactions impose on outside scrutiny. [24][25]

A chart extrapolation discussed in the feed is therefore not a demonstrated deadline for superintelligence. Conditional forecasts about compounding improvements should not be read as observed outcomes. [24][25]

Science and Security: Continuity With Day 23

PSI’s reported AI-assisted Aaronson–Ambainis proof resurfaced in Ryan’s feed. This is the announcement covered in Day 23, not a second discovery: PSI says a system using a non-public Anthropic model produced a Lean-accepted proof. Its quantum-computing implication concerns approximating acceptance probabilities on most random inputs with polynomial query overhead, not a general claim that classical computers efficiently simulate every quantum computation. [26]

Anthropic’s guest article on scientific models also resurfaced through Labs reposts. It describes agents writing, fitting, comparing, and revising models, while emphasizing distributions of plausible explanations and the difficulty of interpretation. Its authors caution that evidence on problems without known answers remains sparse. [27]

The Firecracker escape report is another Day 23 item resurfacing, not a newly verified breach. The report describes host execution on a setup running Firecracker directly as root without its jailer, using an earlier host-memory primitive; production containment requires a separate assessment. It does not by itself show that AI discovered the escape. [28]

A separate interview excerpt quotes Calif’s CEO warning that defenders lack a satisfactory playbook for a thousand agents attacking a network at once. That is a warning about a scenario, not a documented thousand-agent incident or proof that every existing defense fails. [29]

Capital and People: Reported Deals, Not Confirmed Closings

Business Insider reported that Thinking Machines Lab was nearing a $5 billion raise at a $40 billion pre-money valuation, citing people familiar with the matter. Its report is about an expected closing, not evidence that the money had already closed. [31]

Other resurfaced posts asserted an $11 billion OpenAI round at a $1.25 trillion valuation and described declining proposed Anthropic IPO valuations. No primary transaction announcement or filing was established for those claims in this review; they should not be treated as confirmed financing events. [32][33]

Guillermo Flor’s paid newsletter preview describes Martin Casado’s discussion of AI financing and acquisitions, including Cursor and SpaceX. Only the public preview was readable here, not the paid interview breakdown; the deal narrative and training-cost estimates are therefore not independently established in this issue. [34]

A founder-interview account described repeated fundraising rejection before eventual backing, and one former Cursor employee announced a move to Google to work on agentic coding for businesses. The former is a selected anecdotal sample, not a fundraising success rate; the latter is the employee’s own announcement. [30][35]

Physical Systems: Plans Are Not Deployments

Resurfaced posts discussed StarMind orbital computers and projected launches beginning in the third quarter of 2027. The directly observed Musk reply says details of the Starmind computer will be unveiled next year; it does not itself substantiate every launch-date or hardware-detail claim made in the secondary post. No orbital deployment is established here. [36][37]

Musk separately said Tesla would continue building Cybertruck for another eight or nine months before focusing on scaling Cybercab. This is his stated production plan, not a completed factory transition. Other resurfaced physical-world posts concerned AI-guided hardware, modular nuclear-reactor design concepts, and software-driven engineering at Boom; concept imagery and productivity claims should not be confused with deployed reactors, audited engineering benchmarks, or certified aircraft. [38][40]

Coverage note: This issue covers public activity after November 7, 4:00 AM CST through November 8, 5:40 AM CST. The paginated timelines returned 74 Ryan posts—57 quotes, 11 replies, and 6 originals—and 53 Labs posts—52 reposts and the Day 23 article. Both timeline pagination chains reached an end; source posts older than the window are identified as resurfaced context, not fresh events.

One quoted source, post 2119126887330885814, was unavailable. Selected static images were reviewed, including the conditional acceleration chart and a user-report screenshot; embedded article text and available long-note text were read. Videos and the remaining media were not reviewed, so media-only claims and demonstrations remain gaps. Likes were outside scope. Political commentary, social greetings, jokes, and unsupported model-release speculation are recorded in the coverage ledger rather than presented as technological developments.

Sources:

← Back to News