Keep the old reservations.
Agents finish accepted tasks beyond C’s view. Without a completion report, should C give the same task to someone else?
04 / Coordination · Intermediate
Decision
architectures.
Who can decide? Who can know?
The agents keep working. Reports can stop arriving.
Explore how decision rights change when a network splits.
01 / Links cut
2.0 s
The network splits. Accepted work continues.
02 / Work finishes
4.0 s
T3 is finished. Its report cannot reach C.
03 / Reports return
8.0 s
Fresh reports update C. The mission continues.
A3’s new report carries its earlier T3 completion. Dropped packets are not replayed. Drawings are schematic.
A browser simulation of planar work and logical messages. The 3D view adds no radio model, flight physics or live execution.
Builds on Consensus and Task allocation.
Before you begin
No fresh report.
Has the agent stopped?
Compare a central coordinator, delegated groups and peer replicas. Each uses nearest-pair greedy; authority, eligibility and synchronization differ.
Explore the four questionsMethod & assumptionsNearest-pair greedy · declared decision rights · local report caches
0.1 s synchronous rounds · symmetric links · no delay on active links
- Assignment algorithm
- Nearest-pair greedy
Choose the nearest eligible idle-agent / unreserved-task pair, then repeat. No Hungarian solver is used in this comparison.
- Authority
- Central · hierarchical · peer
C assigns globally; A1/A2 assign in fixed domains; or each peer independently agrees on the same plan and takes its own target.
- Peer synchronization
- Full-roster round barrier
All three current reports and matching nonempty plans are required. Missing participants block new peer assignments.
- Execution and transport
- Synchronous · deterministic
0.1 s steps, symmetric links stable during each round, no delay on active links. Cut packets are discarded.
Fidelity: one browser simulates separate local caches and decision rights. Physical 2D/3D views observe the same planar mission. The 3D quadrotors use a fixed display altitude of 1.5 m. Station colors show physical work; observer labels distinguish learned completion and stale reports. The yard is illustrative. No real network, fault detector, flight physics or arbitrary-loss consensus protocol is implemented.
01 / Try it
Work can finish beyond your view.
Select an agent to inspect its cache. ? marks completed work this observer has not learned. Map positions do not determine logical network links.
Physically completed0 / 6Evaluator truth · full service required
Confirmation coverage0 / 6Completed tasks known to C
Simulated time0.0 sStep 0 / 600
Delivered / dropped0 / 0Modeled packets · not network bytes
Inspect received information
C’s report cache
| Agent | Sampled | Age | Reported state | Target | Reported position |
|---|
Watch the difference
Two different measurements
━━ Physically completed┄┄ Confirmation coverage
Confirmations are received knowledge. The evaluator can see finished work before the selected architecture can confirm the whole mission.
Put the views side by side
Compare work with the selected cache.
| Task | Physical state | Executing agent | Service | C knows |
|---|
Inspect actual execution
Three agents remain operational.
| Agent | State | Target | x | y |
|---|
Read the sequence
What changed at the boundary?
Latest 30 events. Reports and agreement phases consume zero model time. Packet counts do not measure latency or throughput.
02 / Follow a question
Keep working.
Keep track of who knows.
Change the decision rights, then inspect what the cut actually prevents.
Continue inside each group.
A1 controls T1/T4. A2 allocates the other four tasks to A2/A3. Their domains survive this cut, but also restrict choices when connected.
Wait for the whole roster.
A peer needs all three current reports and matching plans. Existing work continues; what happens when it needs a new assignment?
Learn what happened out of view.
Reconnect at 8 s. Fresh reports carry earlier completions. Compare what C learns with what allows the peers to assign again.
Each case starts paused. Diagrams explain authority and message delivery, not physical routes. These are specific protocols, not a ranking of all architectures.
Compare all nine reference runs Execution · confirmation · recovery
Same physical mission and reporting pattern. Hierarchical domains restrict eligibility; the peer barrier changes synchronization. These outcomes do not rank every implementation of an architecture.
| Architecture | Network | Physical | Confirmed | All work done at | Final result | Dropped reports |
|---|
03 / Go deeper
Finished work.
Received knowledge.
A controller can only act on what it has learned. A missing report does not cancel an executor’s task or release its reservation.
At 4 s · central coordinator + cut
- Physically finished
- 2 / 6
- Known to C
- 1 / 6
T3 is complete, but C still holds A3’s report from 1.9 s. It says “travelling” and keeps T3 reserved.
Fresh reports at 8 s bring C to 3 known completions. The remaining work still needs execution.
Read the protocol, assumptions & source
- Execute accepted work for 0.1 s.
- Apply the scheduled network state.
- Deliver current reports over active links.
- Compute plans from each local cache.
- Peers exchange and compare nonempty plans.
- Apply permitted commands, then evaluate.
What the groups change
Task domains are assigned before time zero. They are disjoint and never rebalanced. Group performance includes this eligibility constraint; it is not an isolated measurement of hierarchy alone.
What the peers require
A fixed roster, synchronized rounds and symmetric links stable across report and plan phases. There is no leader election, quorum rule or discovery of network components.
What the counters mean
Every agent sends a report to two peers and C per boundary: nine attempted packets. Proposals and remote commands add separate counts; local actions do not.
What the experiment can conclude
Execution plus confirmation completes the run. Otherwise it reaches 60 s, even if the work itself finished earlier. The peer coverage metric is an evaluator summary, not a termination message.
Primary conceptual reference: Leslie Lamport (1978), Time, Clocks, and the Ordering of Events in a Distributed System ↗. This workshop declares its own synchronous replicated-greedy protocol; it does not implement Lamport logical clocks, Raft, Paxos or a general consensus algorithm.
No executor fails in this lesson. Stale reports are not proof of failure. The peer barrier is a teaching protocol for the declared network; these outcomes do not establish a general advantage for any architecture.
Keep the thread · Advanced
Who wins
a disputed task?
Explore how CBBA resolves competing task claims. This separate allocation model builds on Consensus and Task allocation; its agents do not move or execute work.