RELIABILITY FOR SCHEDULED AGENTS

The night shift, documented.

Foreman is the reliability layer for scheduled AI agents. Dedup gates, retries that resume from the last checkpoint, silent-failure detection, and a run report sent wherever you already look.

Built on the Claude API. Tool use · MCP · scheduled runs

06:00Schedule firedok
06:01Dedup gate passedok
06:06Draft written, checkpoint savedok
06:11Render failed, retryingretry
06:13Resumed from checkpointok
06:18Posted to 3 channelsok
gateDedup registry checkpassed
gateCheckpoint writepassed
gateChannel health checkpassed
gateDuplicate topic refusedheld
gatePost landing verifiedpassed
runsCompleted12
retryResumed, not replayed1
heldDuplicates refused1
failSilent failures0
needMessages needing you0
1 retry0 duplicates posted18 min

WHAT IS INSIDE

Four gates between your agent and the publish button.

Deterministic code, not prompt instructions. A gate either passes or the run stops.

gate/dedup

01

Dedup gate

Every planned topic is checked against the publish registry before the agent spends a token. A repeat never leaves the building.

gate/checkpoint

02

Checkpoint resume

Each step writes its output as it finishes. A retry picks up at the failed step instead of replaying the run from the top.

gate/channel

03

Silent-failure detection

After posting, Foreman verifies the post actually landed on every channel. No landing, no silence: you get told, with the channel name.

gate/report

04

Run reports

Every run ends with a short summary sent where you already look. If nothing needed you, it says so in one line and stays out of your way.


THE LOG

One night, three entries.

Scheduling an agent is the easy part. Every failure below happened to a real content fleet, and none of them raised an error.

02:14INCIDENT · RETRY

Render failed at 02:14. The retry did not start over.

Step 4 of 7 died on a transient error. The old way replays steps 1 through 3. Foreman resumed at step 4, because every finished step had already written its checkpoint.

03:40GATE · HELD

The same topic, three days apart. Held at the gate.

The agent planned "Soto Ayam Lamongan" again. The registry showed it ran three days earlier, so the dedup gate refused it before a single token was spent on the repeat. Eyeballing history does not scale. A hard gate does.

05:12CHECK · WARN

Instagram authorization had lapsed. Nothing errored.

Foreman health-checks every channel before the run. At 05:12 it found the Instagram channel down and reported it, instead of letting the 06:00 post silently never go out. The channel was reconnected before breakfast. The audience never noticed.


HOW IT IS STAFFED

Three layers, one job.

The skill gets you started, the MCP server plugs into your agent, the hosted service does the heavy lifting. Only the last one is paid.

Skill FREE

The playbook. Teaches any agent the reliability patterns: gates, checkpoints, idempotent side effects. Distribution, not the product.

MCP server

The sidearm. Your agent calls Foreman tools directly: check the dedup registry, write a checkpoint, file a run report. Works with any agent client that speaks MCP.

Hosted service PAID

The engine. Scheduling, execution, metering, and the report channel. The part you pay for, per run, with nothing to self-host.

mcp-config.json
{
  "mcpServers": {
    "foreman": {
      "url": "https://foreman.grafisify.com/mcp",
      "headers": { "Authorization": "Bearer $FOREMAN_KEY" }
    }
  }
}

GET COVERAGE

First slots open soon.

Put Foreman on your night shift. Leave a work email and we will write when your slot opens. One email, no newsletter.