Agents that have been stuck for more than 5 minutes (or 300 simulated seconds) are listed below. You can hand them off to a human.
How does this dashboard evaluate token costs, track agent execution states, and flag stuck runs for human handoff?
This interactive tool tracks simulated autonomous agent sessions across execution lifecycles (running, completed, stuck, and failed). It tallies combined input and output token counts and calculates expenses using a flat baseline rate of $0.002 per 1,000 tokens ($2.00 per 1M tokens). When an agent's inactivity exceeds a 300-second threshold, the monitor flags it as stuck and queues it in the handoff view, where triggering a handoff transitions its status directly to completed.
The dashboard models token expenditure using a uniform $0.002 per 1,000 tokens for both input and output tokens. Real-world LLM provider APIs implement asymmetric pricing where output tokens cost significantly more than input tokens, alongside separate caching and batch rates. Furthermore, stuck detection relies solely on a 300-second inactivity timer, and clicking handoff immediately marks the job as completed without human review intervention or execution retries.
Click '+ Add Run' on the Runs tab. The monitor appends a newly generated agent run (e.g., RUN-008) with randomized step and token values, and increments the 'Total Runs' counter from 7 to 8 while updating the status breakdown accordingly.
The dashboard calculates run cost with the linear formula ((tokens_in + tokens_out) / 1000) * 0.002, treating prompt input tokens and completion output tokens as identically priced units at $2.00 per 1,000,000 tokens.
In production LLM infrastructure, standard model pricing sharply separates input tokens from generated output tokens, frequently billing output tokens at three to five times higher rates, with distinct pricing tiers for cached inputs, long contexts, and batch processing. Pricing | OpenAI API
The Token Costs tab initializes an HTML5 canvas element managed by Chart.js to render a responsive bar chart mapping run identifiers to individual run cost estimates in USD. Getting Started | Chart.js
When switching between tabs or dynamically adding runs, existing chart instances are destroyed and re-instantiated so that updated token sums reflect in the canvas scales without memory leaks or overlapping graphic buffers. Getting Started | Chart.js
Commercial API token fee structures differentiating input, cached input, and output token costs across model families.
Canvas-based dynamic bar chart creation and responsive chart updates on HTML web pages.
The monitor starts with seven fixture records: two completed, two running, two stuck and one failed. Text filtering narrows visible rows but summary statistics use the entire array. The handoff list filters directly on stored status equal to stuck. Despite the five minute description and periodic refresh comment, no three hundred second threshold converts a running record into stuck in the executable render function. This is local demo data, not active agent telemetry. The fixture contains twenty eight thousand seven hundred fifty total input and output tokens. Its independently stored costs sum to point one one eight dollars. Applying the rate used for newly added runs, point zero zero two dollars per thousand tokens, would give point zero five seven five. The cost summary adds stored cost fields; it does not recompute seed costs from token counts. That rate is a simulator assumption, not verified current provider pricing. Handoff to Human sets the selected local record status to completed, refreshes its timestamp, shows an alert and rerenders. It does not send a human a message or execute the remaining work. Added runs use random statuses and counts. Share links serialize the current records; imported links overwrite activity timestamps according to status. The chart and records are illustrations. No provider, production agent or external handoff was used to author or verify this explanation.