It Looked Finished on Day One
I wanted a team of AI agents that could answer a research question the way a small team of people would. One plans the work, one searches the web and reads the sources, one writes it up, and one checks the write-up before I see it. I also wanted to watch them do it: which agent is working, what it's reading, what it handed to whom, and what it cost. The app is called FORGE, and it runs on a small…
The FORGE app, a dashboard for AI agents, was initially a simulation with no real agents. This early version featured a workflow canvas, live event stream, timeline, replay scrubber, agent counters, and cost monitors. Behind the scenes, a simulator produced fake events to display on all screens. One crucial design decision was that each screen was computed from a log of events, not an updating record.
When real agents replaced the simulator, the screens remained unchanged, as they all displayed the same events. The app began with three agents: a researcher, a writer, and a reviewer. The researcher used the GLM-5.3 Flash model from OpenRouter, costing $0.15 per million input tokens. The writer and reviewer used free tools like DuckDuckGo and Claude.
During initial testing, calls to the models frequently timed out due to Node.js's default 250-millisecond attempt limit. The fix involved preferring IPv4 addresses and extending the attempt time to three seconds. Some model replies were empty because the model had exhausted its output budget. The researcher continuously researched until it hit the step limit, resulting in incomplete notes. Each tool result also displayed the remaining rounds, forcing the researcher to write up its findings once the limit was reached.
Real agent implementation revealed additional issues. The run history contained sample runs, not actual data, creating confusion. Simulated data needed clear labeling to differentiate it from real data. The server now maintains a single SQLite database, eliminating per-browser local storage conflicts, and issues unique run IDs to prevent data overwriting. The reviewer proved to be the most expensive component, with Claude Sonnet offering the best balance of accuracy and cost.
Written by urgent.news from Dev.to's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.
