ROUTE · INSPECT · COMPARE
Agent market: assign work under skill and budget constraints
Inspect eligible bids, award costs, and tasks left unassigned by a tight cap.
Four task decisions · sixty-second inspection0 / 60 seconds
Task trace
Award policy and scope
This local teaching policy awards independent tasks; it does not execute them. The original Contract Net paper distinguishes eligibility, offers, and awards. The lowest-eligible-price rule here is an illustrative choice, not an ADK implementation.
Separate an award from completed work
This local market allocates four independent teaching tasks: retrieve evidence, check a claim, write an answer, and approve release. It does not execute any task. An award commits the winning bid price in the fixture; it provides no evidence that the bidder produced a correct answer. Although the task names suggest a workflow, this experiment does not enforce dependencies between them. The complete allocation is computed immediately. Playback highlights each decision for fifteen seconds so you can inspect its inputs; the summary explicitly distinguishes computed awards from decisions already inspected.
Inspect every offer
There are two fixed bidders for each task. The cheap bidder offers three, four, two, and five credits respectively. The premium bidder offers one additional credit for each task. Both are qualified except that the cheap bidder lacks approval skill. Every card and the text trace display each offer and all applicable rejection reasons: unavailable, missing approval skill, or over budget. The award goes to the lowest available, qualified offer within the per-task cap. Equal prices would use bidder identifier order; this fixture has no ties. Eligibility cannot be bought by increasing the cap.
Change the allocation
Budget per task selects a three-credit or six-credit ceiling. With the low ceiling and both bidders available, retrieval and writing receive awards totaling five credits; checking and approval remain unassigned. Switch to the high ceiling: all four receive awards totaling fifteen credits, including six credits to the premium approver. Unavailable bidder removes the selected bidder from consideration for every task. At the high ceiling, removing the cheap bidder yields four premium awards totaling eighteen credits. Removing the premium bidder leaves three cheap awards totaling nine credits and no qualified approver. At the low ceiling with cheap unavailable, only premium writing fits: one award for three credits.
Control inspection time
Play sequence starts the sixty-second inspection and changes to Pause sequence. Pause settles elapsed time and freezes the highlighted decision. Restart pauses and returns to zero without altering the chosen budget or availability. Timeline seeks to any inspection second and pauses. Speed selects half, normal, or double pace; changing it during playback first accounts for elapsed time at the previous speed. Changing either allocation input starts a new paused comparison. A hidden tab pauses rather than silently advancing. These are inspection seconds, not measured agent latency, cost accrual, or execution time.
Recognize the policy limits
The cap applies separately to each task, not to a shared project budget. The fixture assumes unlimited bidder capacity, fixed truthful prices, no failures after award, and no strategic bidding, retry, reputation, fairness or settlement. Eight modeled solicitations means one request to each of two bidders for each of four tasks, including unavailable bidders; it excludes replies and awards and is not a network packet measurement. Contract Net motivates separating eligibility, offers and awards; this cheapest-eligible rule is a chosen local policy, not a complete implementation of that protocol or ADK Java.
Keep a comparison
Capture baseline saves visible outcomes, Replace baseline updates the snapshot, and Clear baseline removes it. Notebook text is stored locally; Export notes downloads those observations. Expand experiment enters fullscreen and Exit full screen returns to the page. Use the two related readings for broader agent coordination context. Compare the full bid trace as well as totals: an unchanged total can hide different ownership, and an unassigned approval is a material limitation even when other tasks have awards.