Agent harness for multiple businesses

AI operations, multi-venture

The client ran two businesses, DeFi growth and B2B services, on an open-source AI agent. We rebuilt that deployment into a production operations layer: high-volume automation with human-gated execution end to end, run by a single operator.

Key figures

• 3.7M: tokens a day recovered from idle polling, an approximate figure
• 892: lifecycle sends a day re-lit within 48 hours
• Under 3: notifications a day, down from 18 to 30
• 6,619: upstream commits rebased behind a test gate

The challenge

One operator, two businesses, and an automation layer grown by accretion on a fast-moving open-source agent framework. The agent burned millions of tokens a day polling channels that produced nothing; dashboards showed 46 of 46 green while zero emails actually sent; six lifecycle email journeys sat dark for a week; the fork had drifted 6,600 commits behind upstream; and at 18 to 30 notifications a day, the operator had stopped reading any of them.

The insight

Every real defect shared one shape: components reported on themselves, not on their output. A green scheduler says the job ran; it says nothing about whether the email sent, the post published, or the money moved. Once health was redefined as verified business outcomes, and every irreversible action routed through a one-tap human gate, the machine could be trusted with volume while the operator kept judgement.

What we did

1. Audited every scheduled job against its actual output, and replaced always-on model pollers with deterministic gates that only wake the model when there is real work.
2. Replaced self-reported health with outcome contracts and canaries: checks that fail when the business result is missing, not when a process is.
3. Made upstream adoption routine with a twelve-test safety net that fails on stock upstream, turning thousand-commit rebases into a scripted, reversible operation the client runs themselves.
4. Consolidated 18 to 30 daily pings into one actionable feed with a closed catalogue of one-tap approvals; nothing publishes, sends or spends without the operator's tap.
5. Instrumented revenue truth end to end: person-level email-to-deposit attribution, measured sequence reply rates, and a daily call sheet sized to real capacity.

Outcomes

• Agent token burn from idle polling: Eliminated: about 3.7M tokens a day recovered
• Dormant lifecycle emails: Re-lit: 892 sends a day within 48 hours
• Operator notification load: 18 to 30 a day down to under 3, all actionable
• Upstream drift: 6,619 commits rebased behind a test gate; updates now routine
• False-green monitoring: Eliminated; health now means the output happened
• Qualified lead pool: Four times the pool in one gated run, 130 to 526 enriched leads

What changed

The harness now runs both of the client's businesses daily: content pipelines with validator gates, lifecycle email ladders under ramp caps, lead scoring on live hiring signals, and a morning call sheet that puts their operator on the phone with the twenty right people. Every send, post and spend still terminates in a human tap; the machine builds the lists, the operator does the talking.

Health is measured at the output: if the sends do not happen, the client knows the same day, from one feed rather than thirty pings. The safety net and the update path were handed over as scripted, documented operations the client runs without us.

Evidence

Services: AI operations audit, Agent harness architecture, Workflow automation, Approval and safety design, Growth instrumentation.

The client is a founder-led venture portfolio. All six outcomes are verified.