Supervising AI Agent Fleets: Routines, Watchdogs and Shared Memory in Daily Operations

A single AI agent completing a single task is now an ordinary event. The harder problem, and the one that determines whether autonomous software is useful in a real business, is what happens when many agents run every day, on schedules, touching different systems, without a human watching each step. Luis Lacerda, an AI systems architect working at the intersection of LegalTech and lead generation compliance, treats that problem as an operational discipline rather than a prompting exercise. He designs and operates a fleet of more than ten coordinated agents covering coding, code audit, recruiting, outreach, team-chat monitoring, search indexing and media production. This essay describes the supervisory principles behind that fleet: scheduled routines, watchdogs, and shared, versioned memory.

From Individual Agents to an Operating Fleet

The conceptual shift Lacerda emphasizes is that a fleet is not a larger agent. It is closer to a small organization. Each agent has a defined remit, a defined set of permissions, and a defined output that someone, human or machine, will check. A coding agent proposes changes; an audit agent reviews code it did not write; an indexing agent reports what it submitted; a monitoring agent summarizes conversations rather than acting on them. Separation of duties, a familiar idea in finance and security, becomes the first structural safeguard.

This framing matters because failure in a fleet is rarely dramatic. It is usually quiet: a routine that stopped running, a task that reported success without producing evidence, two agents working from different assumptions about the same fact. Supervision exists to make those quiet failures visible.

Routines: Making Autonomy Predictable

Lacerda's agents run on scheduled routines rather than ad hoc invocations wherever the work is recurring. A routine has a stated purpose, a cadence, an expected artifact and an owner. That structure converts autonomy into something predictable. When an indexing routine runs each morning, the question is not whether the agent "felt" productive but whether the expected report exists and contains verifiable entries.

Routines also impose scope. An agent asked to perform one recurring job, with clear inputs and outputs, is far easier to evaluate than an agent given open-ended instructions. In Lacerda's practice, the specification of a routine is written before the routine is enabled, consistent with the principle he applies across his work: governance before generation.

Watchdogs: Supervision That Does Not Sleep

Routines answer the question of what should happen. Watchdogs answer the question of whether it did. Lacerda pairs critical routines with watchdogs whose only job is to observe: confirming that a process is alive, that a scheduled task produced its artifact, that a long-running job is still progressing rather than silently stalled.

The design choice here is deliberate modesty. A watchdog that tries to repair everything becomes another source of unsupervised change. Lacerda's preference is for watchdogs that detect, record and escalate, with automatic recovery limited to well-understood, reversible actions such as restarting a stopped process. Anything ambiguous is surfaced to a human with the evidence attached. This mirrors how mature operations teams treat alerting: the goal is a short path from anomaly to informed decision, not an autonomous system that hides its own mistakes.

Shared, Versioned Memory: One Source of Operational Truth

The most distinctive element of the fleet is its shared memory. Agents that each maintain private notes inevitably drift apart; one learns that a procedure changed, the others do not. Lacerda addresses this with a shared memory repository used across agents, versioned so that every change to collective knowledge has a history.

Versioning provides three properties that ordinary note-taking does not. First, provenance: it is possible to see when a fact entered memory and which task produced it. Second, reversibility: an incorrect lesson can be corrected without losing the record that it once existed. Third, consistency: agents retrieve from the same source, so a correction made once propagates to all of them. In effect, memory is treated with the same rigor as code, reviewed, versioned and auditable.

This approach also supports retrieval-augmented work. When an agent begins a task, it can consult what the fleet already knows about the system, the client constraints and previous failures, rather than rediscovering them at cost.

Evidence Before Completion

Across routines, watchdogs and memory, one rule holds: a task is not complete because an agent says it is. Lacerda requires evidence proportionate to the claim. A deployment is complete when the live page responds correctly; a code change is complete when tests pass and review approves it; an outreach batch is complete when the record shows what was sent and why it was permitted. This discipline connects directly to his broader work directing AI coding agents under written specifications and enforcing quality through more than 1,000 API tests.

In regulated domains the principle carries additional weight. Legal lead generation is governed by TCPA, CAN-SPAM and CASL obligations, and any agent that touches outreach or consumer data must operate inside those constraints. Supervision is therefore not only an engineering preference but a compliance requirement: the organization must be able to show what its automated systems did.

Human Judgment as the Final Layer

None of this removes the human. Lacerda's model places human judgment at the points where it adds the most value: setting specifications, approving consequential changes, deciding ambiguous escalations and revising the rules the fleet follows. Agents handle volume and repetition; the architect handles intent and accountability. The result is a fleet that scales the amount of careful work performed each day without scaling the amount of unexamined risk.

Practical Lessons

• Give every agent a narrow remit and a verifiable output; separation of duties applies to software workers too.
• Write the routine's specification before enabling it, including the artifact that proves it ran.
• Keep watchdogs observant and conservative; detect and escalate before attempting repair.
• Treat shared memory like code: versioned, reviewable and correctable.
• Define completion by evidence, never by an agent's self-report.
• In regulated work, design supervision so the record can answer a compliance question later.

Conclusion

The value of AI agents in daily operations depends less on any single model's capability than on the structure around it. Luis Lacerda's approach, built on scheduled routines, modest watchdogs, shared versioned memory and evidence-based completion, turns a collection of autonomous tools into an operating fleet that can be trusted, inspected and improved. It is the same conviction that runs through his LegalTech and compliance work: governance before generation, evidence before completion, and measurement before scale.