Cost-Aware Automation

What it is

A workflow design pattern that avoids unnecessary model calls with thresholds, no-model checks, and narrow prompts.

Operating shape

• Difficulty: advanced
• Cadence: continuous
• Category: operations

Core concepts

• Cost Control
• Wake Conditions
• No-Model Jobs
• Monitoring Jobs
• Provider Switching

Source evidence

• vendor/hermes-agent/toolsets.py#<file> — Structural Source Map fallback for toolsets.py. A core source file in the Tool Registry and Toolsets structure within Tools, Toolsets, and MCP.

Related articles

• Seven Cost-Aware Hermes Workflows That Go Beyond Chat
• How to Use Hermes Better Than Most People: Architecture, Project Isolation, and Scalable Workflows

Implementation pattern

Outcome

Run a deterministic threshold monitor in cron’s no-agent mode so normal ticks are silent, threshold ticks deliver a literal script message, and no agent or model is invoked on any tick.

Before you start

• Use a disposable ~/.hermes/scripts test script and local delivery first.
• Choose a numeric metric that a shell script can decide without interpretation. The fixture uses metric.txt and threshold 85.
• Do not add a prompt or skill: this workflow is intentionally script-only, no-agent, and no-model.
• The built-in scheduler runs inside the Gateway. Before creating the job, hermes cron status must report a running Gateway and a healthy, recent ticker heartbeat; stop if it reports stopped, stalled, or failing ticks.

Input and output contract

Input is one integer in metric.txt. For values below 85, the script exits 0 with empty stdout and delivery is silent. For values 85 or above, stdout is exactly ALERT metric=<value> threshold=85. Invalid input exits non-zero so cron records an error. The script makes no network call and invokes no model.

metric.txt = 42  -> empty stdout, no delivery, zero model calls
metric.txt = 91  -> ALERT metric=91 threshold=85, zero model calls
metric.txt = bad -> non-zero error, zero model calls

Build it

Verify the installation and scheduler first. Then create the disposable input and reviewed script and schedule it with the documented no-agent flag. The fixed path keeps the scheduled run independent of the directory from which the gateway was started.

hermes doctor
hermes gateway status
hermes cron status

mkdir -p "$HOME/hermes-labs/cost-aware-automation" "$HOME/.hermes/scripts"
printf '42\n' > "$HOME/hermes-labs/cost-aware-automation/metric.txt"

cat > "$HOME/.hermes/scripts/metric-watchdog.sh" <<'EOF'
#!/usr/bin/env bash
set -eu
VALUE=$(tr -d '[:space:]' < "$HOME/hermes-labs/cost-aware-automation/metric.txt")
case "$VALUE" in (*[!0-9]*|'') echo "invalid metric" >&2; exit 2;; esac
if [ "$VALUE" -ge 85 ]; then
  printf 'ALERT metric=%s threshold=85\n' "$VALUE"
fi
EOF

chmod +x "$HOME/.hermes/scripts/metric-watchdog.sh"
hermes cron create "every 15m" \
  --no-agent \
  --script metric-watchdog.sh \
  --deliver local \
  --name "metric-watchdog"
hermes cron list

Verify it

• Put 42 in the fixture, run hermes cron run metric-watchdog, and confirm the tick is silent.
• Put 91 in the fixture and run again; local cron output is exactly the expected alert.
• Put bad in the fixture; the run records an error rather than a false alert or silent success.
• Run hermes cron list; its entry must show Script: metric-watchdog.sh and Mode: no-agent (script stdout delivered directly), with no skills listed. These observable fields are the acceptance evidence that Hermes selected the script-only path.
• Provider dashboards are optional supporting evidence only: delayed or aggregated counters cannot prove one tick made no model call. The job’s no-agent mode plus its exact script stdout/error artifacts are the required evidence.
• Finish with hermes cron pause metric-watchdog, then hermes cron list --all; leave the lab job paused.

If it fails

• If a model call occurs, pause the job, remove it, and recreate it with --no-agent --script; do not patch around a prompt-driven job.
• If low values send messages, ensure the script emits absolutely no stdout on the normal path.
• If errors disappear silently, preserve the non-zero exit and inspect cron output/logs rather than converting failures to exit 0.
• Use hermes cron pause metric-watchdog while repairing. Rerun all three fixtures, then pause it again; do not leave this lab job active.

Safety, privacy, and cost

The fixture is non-sensitive and the script does not use the network. Do not echo tokens, customer data, or full environment variables into stdout because delivered output is retained. No-agent mode pipes deterministic script stdout directly to delivery and never touches a model, so each tick has zero model cost; host compute and any external message-provider fees can still exist. Keep local delivery until the alert is proven.

Official references

• Script-only cron jobs
• Cron jobs
• Cron troubleshooting