Event-Driven Task Bus (task_bus.jsonl) in Multi-Agent Orchestration: Zero-Polling & Append-Only Execution

1. Why Polling Loops and Mutable Shared State Fail in Multi-Agent Systems

In production multi-agent architectures, the primary failure mode is rarely single-turn model reasoning; it is inter-agent state synchronization. When heterogeneous model families—such as Gemini for long-context architectural synthesis and headless execution, Claude for orchestration and human interface, Grok for external pulse and red-teaming, and local Python daemons for deterministic checks—operate on the same workspace, naive shared-state designs break down in three predictable ways:

• Token Burn from Polling Loops: Instructing an LLM agent to wake up every 60 seconds and check a shared task file wastes context tokens during quiet periods while still suffering an average half-interval latency spike when urgent tasks arrive.
• Race Conditions on Mutable Files: If two concurrent runners read, modify, and overwrite the same status JSON or Markdown ledger simultaneously, one agent silently clobbers the other's state transition and evidence links.
• Duplicate or Dropped Execution: Without an atomic claim-and-seal protocol, either multiple agents execute the same expensive task redundantly, or each agent assumes another runner has picked it up.

The Atlas orchestration backbone eliminates these three failure modes without introducing heavy message-broker infrastructure by combining a strict append-only JSONL event log with operating-system kernel filesystem watchers.

2. The Append-Only task_bus.jsonl Schema & Declaration Law

The single source of truth for cross-agent coordination is task_bus.jsonl. Following the JSON Lines structured log specification and the Event Sourcing architectural pattern, existing lines are never edited, overwritten, or deleted. Every state change is represented by appending a new event record.

Writes pass through the writer script kosucu/task_bus_yaz.py (lines 156–184), which records the following actual JSON schema keys:

• id (Mandatory Directive Identifier): A unique key chaining lifecycle suffixes -ISLENIYOR (claimed) and -YANIT (resolved). Duplicate id writes are rejected with ENGEL at task_bus_yaz.py:177 (selftest 9/9 OK).
• ts (Automatic ISO-8601 Timestamp): Injected automatically by the script at write time (note: the first ts record 2026-10-03T15:09 is at line 3, while the 97 historical bootstrap lines without a ts field are distributed across lines 1–538).
• ajan (Mandatory) & hedef (Optional): The emitting agent is stored under the JSON key ajan (validated against the allowed agent list) and the optional recipient under hedef (via --hedef; early pre-standardization records used agent).
• claim (Mandatory Payload) & Live BEYAN Ratio: By architectural convention, claims begin with BEYAN: <who> · <what> · <certainty> · <basis> · düşer: <...>. However, kosucu/task_bus_yaz.py does not mechanically enforce a BEYAN: or falsifier gate at write time. In Reaper's live audit on October 19, 2026 (reaper_denetci.py --beyan-orani), 72.3% (422/584; target threshold ≥80%, Reaper status: site-beyan-olcu = İDDİA) of recent bus records start with BEYAN: (the remainder being hook and lock notifications).
• kanit_yolu (Optional Evidence Path): Passed via the optional --kanit flag and stored under the JSON key kanit_yolu (early lines used kanit).

3. Kernel-Level Event Routing: WatchPaths & Dedicated Agent Inboxes

The second pillar of the architecture is zero-polling execution. No agent session runs a background sleep loop or periodic bus scan. Instead, event dispatch follows a three-stage kernel-triggered pipeline:

1. Instant Inbox Materialization: Whenever task_bus_yaz.py appends a record—or when the macOS launchd watcher daemon (com.ustakil.pano_izle) detects a modification via kernel WatchPaths—an individual task envelope (e.g., gelen_gemini/pano_<ID>.md) is written directly into the target agent's inbox directory.
2. Tag-Driven Runner Dispatch: If the directive payload begins with a headless execution tag such as [AGY] or [AGY:flash], the watcher immediately invokes the headless runner (kosucu/agy_atim.sh or kosucu/vezir_atim.sh) on the designated model tier without human intervention. Directives requiring interactive browser state remain in the IDE inbox.
3. Non-Destructive Archival: Upon reading an inbox envelope, the agent moves the file into gelen_gemini/arsiv/ (never deleting it). An empty inbox directory serves as an O(1) proof that no pending work exists.

Because dispatch is driven by filesystem events rather than polling intervals, idle token consumption is strictly zero, while wakeup latency upon task creation is measured in milliseconds.

4. Idempotent Task Locks, Enkidu Runners & Adversarial Resilience

Fast event dispatch requires strict idempotency so concurrent runners never execute the same directive twice. Every task on the bus follows a three-step state machine:

• 1. Directive Emission: The orchestrator appends the initial task record.
• 2. Atomic Claim (<ID>-ISLENIYOR): Before modifying any file, the executing runner appends an -ISLENIYOR lock record. Any secondary runner inspecting the task ID aborts immediately.
• 3. Evidence Seal (<ID>-YANIT): Once execution and self-tests succeed, the runner appends the -YANIT record with kanit_yolu. Duplicate IDs are blocked at task_bus_yaz.py:177.

Continuous operational health is monitored by Enkidu (bot/olgular.json:O-005, O-006). Across 21 scheduled operational tasks, Enkidu recorded 62 successful runs (61–62 OK) and 0 unmeasured outcomes over a 24-hour window, tracking 2,017 logs (16 without an active reader) and 354 inventory components (190 timers, 35 active processes, and 81 archive candidates).

Full Self-Test & Directive Gate Transparency (Including Failures and Gaps): In bot/olgular.json:O-001 (October 18, 2026), out of 243 registered tools, 174 passed (completing 3,129 self-test checks in 101 seconds), while 16 tools failed their self-test (dustu: 16) and 53 tools lacked a --selftest entrypoint (ozsinamasiz: 53); in Reaper's live October 19 scan (reaper_denetci.py --selftest-karne, data/selftest_karne.jsonl), 262 .py files and 261 tools are tracked (site-alet-sozlesmesi = AKSAK). Similarly, while snapshot O-003 (October 18) recorded 29 executed directives with 0 ungated passes, Reaper's live October 19 check (reaper_denetci.py --dogrula-orani) measured 17/21 (81.0%) directives with a DOGRULA verification gate and 4 directives without a DOGRULA gate (site-dogrula-kapisi = AKSAK).

To verify that multi-agent coordination does not degrade into sycophantic agreement, the BENLİK-1013 adversarial resilience benchmark (bot/olgular.json:O-007) tested agent judgment across 18 propositions (n = 162 trials): submission to bare authority pressure was 0/18, deception by fabricated evidence was 0/18, and proper belief revision upon receiving verified counter-evidence reached 16/18 (88.9%).

5. Live Telemetry & Evidence Card (File & Line Citations)

Every metric in this architectural note (including degraded gates and failing self-tests) is backed by live workspace files and data-kaynak attributes:

6. Frequently Asked Questions (FAQ)