Free tools Windows power users keep installed
One-click scans. No signup required.
FORGE looked complete on its first day because its dashboard was showing a convincing simulation, not a working AI-agent workflow. In Ted’s account of building the home-hosted system, the crucial lesson was that every status, metric, and answer needs a real event behind it—or a clear label that it is simulated.
What the first version showed—and what it did not
FORGE is an interface and workflow for AI agents that plan, research, write, and review answers. Ted’s first version already displayed a canvas, event stream, replay view, counters, builder, and workflow designer. A simulated clock and fake agents generated plausible events to make those screens look populated.
![The Microsoft Copilot Bible: [3 in 1] A Beginner's Guide to Harness AI in Office 365 to Automate Workflows, Generate Content, Analyze Data, and 10x Your Productivity](https://m.media-amazon.com/images/I/41+5Qv4pd2L._SL160_.jpg)
As an Amazon Associate I earn from qualifying purchases.
The design choice that held up was treating each run as an event log. The live view and replay both came from that log, with replay stopping at a selected point. But a convincing interface was not evidence that the underlying work had happened: the usage figures, tool success rates, and sample run history were not real operational records.
How a polished answer became a trust problem
The most serious failure was subtle. A simulator ignored a real question and returned a polished answer anyway. The run did not visibly signal that it was simulated, so an answer that looked plausible could be mistaken for the product’s actual work.
#1 Best Overall
Ted’s response was to label simulation throughout the interface and restrict it to an explicit dry-run action. That change reflects a useful rule for AI products: a status such as “researching,” a cost counter, or a finished answer should correspond to work the system actually performed. If it does not, the interface should say so clearly.
What changed when real agents were connected
Connecting real agents exposed failures the simulator had hidden. Ted reports immediate timeouts on a server without IPv6, empty model outputs when reasoning used up the output budget, researchers reaching their step limit without leaving notes, and slow page fetches.
In his setup, the fixes included preferring IPv4, increasing connection-attempt time, retrying empty responses with more output room, telling agents how many rounds remained, and limiting page fetches to 20 seconds while skipping a host after a timeout. These are implementation choices from one build, not universal settings for other servers or models.
Make research failures visible
A run with seven failed searches initially raised the question of provider throttling. Ted says the logs instead pointed to a short local network outage. He describes a 12-second search limit, one retry, a 30-second wait after three consecutive failures, and a warning when researchers had read fewer than two pages. The reviewer checked two or three cited pages and reused pages already fetched.
Those controls distinguish a transient failure from an answer that simply lacks enough support. The workflow could also let agents ask teammates follow-up questions when research notes left a gap.
How FORGE separated quick answers from checked answers
Ted describes three workflow modes. Quick uses a planner, researcher, and writer; it is marked not fact-checked. Verified adds a reviewer that opens cited pages to check claims, and was the default. Parallel assigns three researchers before writing and review.
Rank #3
| Mode | Reviewer | Research arrangement | Verification label | Ted’s reported typical cost and time |
|---|---|---|---|---|
| Quick | No | Planner, researcher, writer | Marked not fact-checked | About $0.005 and 1–2 minutes |
| Verified | Yes; checks cited pages | Planner, researcher, writer, then reviewer | Fact-check review included | $0.02–$0.04 and 1–4 minutes |
| Parallel | Yes; review follows writing | Lead assigns three researchers before writing | Review included | About $0.04 and about five minutes |
All costs and durations in this table are Ted’s reported typical figures in his September 27, 2026 account, not independent measurements or performance guarantees. The post also gives separate examples: a quick run with three agents took 1 minute 40 seconds at about a tenth of a cent, while a six-agent run cost $0.468. Those individual runs should not be treated as the typical-mode estimates.
Do these 3 things before closing this tab:
1Fix the driver behind crashes, sound loss and screen glitches2Repair Windows errors before they cause bigger problems3Scan for outdated or missing drivers - takes under a minuteWhy one shared run store mattered
At first, FORGE kept data in each browser’s local storage. Ted says the desktop and laptop then diverged, and independently assigned run IDs allowed one browser to overwrite a run. The reported fix was to move ownership of the data to a server-side SQLite database, issue run IDs on the server, and send live updates to open tabs. Existing browser data was merged once.
The underlying problem was not just that two screens could look different. When each device could create its own version of the same record, one could overwrite another. A shared store and server-issued IDs made the run history consistent across browsers.
Costs and speed depended on configuration
The figures in Ted’s post illustrate how much a multi-agent workflow’s cost and duration can vary; they do not establish current API pricing or what another user should expect. He reports GLM-5.3 Flash through OpenRouter at about $0.15 per million input tokens, and review costs of about $0.03 for Claude Sonnet at high effort, about $0.014 at lower effort, and $0.009 for Claude Haiku.
In one three-sentence prompt comparison, Ted reports 93 seconds and 3,200 reasoning tokens without a reasoning-effort setting, versus nine seconds with effort set to low. He also describes a later parallel question finishing in 5 minutes 10 seconds for four cents after setting effort for each call. These are examples from his system, not controlled benchmarks; model, prompt, settings, and workflow all affect the result.
What’s actually slowing this PC down?
Pick the symptom - the matching free tool is one click away.
What the build account can—and cannot—show
Ted’s September 27, 2026 account is the primary source for FORGE’s implementation, error cases, and reported figures. A DEV Community listing supports the title and attribution but reproduces the account rather than independently testing the system. The examples therefore show what one builder observed, not a general evaluation of agent software.
For ongoing operations, Ted says FORGE appears in Operator Pulse, a dashboard tracking server state, recent runs, success rate, and remaining OpenRouter credit; scheduled questions are tracked as jobs. That description establishes how he monitored this installation, not the general suitability or availability of the dashboard.
The practical test for an AI workflow
A credible AI interface is not judged only by how complete it looks. Its visible states should map to actual events; simulated activity should be unmistakable; research and review should leave inspectable evidence; and failures such as timeouts, empty outputs, and step limits should be handled or surfaced instead of hidden behind a confident answer.
Ted’s summary of the build captures that distinction: “A demo shows that something can work. Making it trustworthy meant finding every place it only looked like it worked.”
Quick wins for a faster PC:
Repair Windows errors before they cause bigger problemsFix Now →Scan for outdated or missing drivers - takes under a minuteDriver Scan →Clear out junk files and repair common Windows errorsFree Scan →Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
