Build agents and workflows

Use ParticleDB for an agent’s records, retrieved context and execution state. The examples here describe application-managed workflows. They do not require a built-in agent scheduler.

Prepare context

Retrieve relevant records with search and memory. Scope data access to the authenticated user or account, and pass only the context required by the next step.

Store execution state

CREATE TABLE agent_runs (
  id BIGINT PRIMARY KEY,
  status TEXT NOT NULL,
  attempts BIGINT NOT NULL,
  result TEXT
);
INSERT INTO agent_runs VALUES (1, 'pending', 0, NULL);

Record a completed step

After your worker finishes the work, record its result in a transaction:

BEGIN;
UPDATE agent_runs
SET status = 'completed', attempts = attempts + 1, result = 'reviewed'
WHERE id = 1;
COMMIT;

SELECT status, attempts, result FROM agent_runs WHERE id = 1;

The result is completed, one attempt and reviewed.

Handle retries

Give each logical action a stable identifier. Retry database transactions after a retryable conflict. For external actions, use the external system’s idempotency support or record enough information to reconcile the result before trying again.

Database recovery does not guarantee exactly-once delivery of a payment, message or model request. Workers must handle the gap between an external response and a database commit.

Put reusable logic in Particles

Use Particles for supported scoring, transformation and computation. The availability page explains which execution models are supported.

Related

Documentation home · Troubleshooting