Until today, Oceum's agents were reactive. A cron fires, the agent runs, it reports what it found, then it sleeps. For fifteen minutes or six hours or until Monday morning, the agent is blind. Things happen in between. Patterns emerge across domains. Nobody notices until the next scheduled run.
Version 2.3 changes that. The Autonomous Pulse System gives the Oceum agent estate persistent awareness — a heartbeat that runs every five minutes, synthesizing observations from every agent, deciding whether to act or stay quiet, and interrupting humans only when the reasoning justifies it.
The problem with scheduled agents
A security agent that runs daily at 6 AM catches yesterday's anomalies twelve hours late. An uptime monitor that checks every fifteen minutes misses the correlation between a latency spike and the payment failures happening in a different agent's domain. A triage agent that processes tickets every two hours doesn't notice that three tickets in the last thirty minutes all reference the same broken endpoint.
Individual agents are domain-narrow. The gaps between their scheduled runs are operationally invisible. The platform has knowledge that no single agent possesses — but only if something is watching the whole picture.
How the Pulse works
Observation layer. Every agent cron now records what it notices — not just what it acts on. The security agent records unusual auth patterns. The uptime monitor records latency thresholds. The triage agent records ticket surge indicators. These observations are cheap database writes with no LLM calls. Eleven agent crons, each contributing zero to three observations per run.
Situation snapshot. Every five minutes, the Pulse cron builds a fast, read-only snapshot of platform state: recent errors, new memories, agent status changes, pending observations. It hashes the snapshot. If the hash matches the previous tick, nothing changed — the Pulse skips the LLM call entirely. In steady state, this eliminates roughly ninety percent of unnecessary API calls.
Act-or-quiet decision. When the snapshot has changed, the Pulse sends a single, stripped-down prompt to Claude Haiku. Not a full reasoning pipeline — a focused binary question: is there something here worth acting on? The model sees all pending observations across all agents and decides. Every decision is recorded in an append-only journal with full reasoning.
Confidence-gated notifications. When the Pulse decides to act, it doesn't blindly alert. Notifications flow through a two-tier gate. Immediate alerts fire only when confidence exceeds ninety-five percent and severity is critical — maximum three per day. Everything else is batched into a daily digest delivered at 8 AM. Per-topic cooldowns prevent the same alert from firing repeatedly. Every notification includes why the agent decided to interrupt.
Overnight consolidation
At 3:30 AM, a separate agent — the Consolidator — runs a four-phase memory cycle we call Reverie. It orients itself in the current memory landscape, gathers the day's observations and journal entries, consolidates related memories by resolving contradictions and merging duplicates, then prunes what's no longer relevant. Proposed changes go to a staging area before promotion. An advisory lock prevents concurrent modification. The Consolidator can never erase its own history.
The result: when agents wake up the next morning, their shared memory is clean, compressed, and contradiction-free. No unbounded accumulation. No stale context degrading decision quality over time.
Governance built in
The agent journal is append-only by design. No UPDATE. No DELETE. Monthly range partitioning keeps query performance stable as the journal grows. Every observation, every decision, every action, every deferral is traceable. Enterprise auditors can reconstruct the complete chain: what the platform noticed, how it reasoned, what it chose to do, and what it chose not to do.
This is not observability bolted on after the fact. The governance layer is the architecture.
What's next
The Pulse System ships with fixed five-minute intervals and calibrated thresholds. Future versions will adapt tick frequency based on platform activity, add a dashboard for journal transparency, and extend notification channels beyond Telegram. The foundation is in place. The agents are aware. They never sleep.
The Autonomous Pulse System is live in Oceum v2.3. Five phases, ten plans, twenty-four files, twenty-seven hundred lines of code. Your agents now notice, reason, and decide — even between their scheduled runs.