TL;DR: What once required three to four people and two to three weeks of setup time can now be managed by a single operator in an afternoon. Eric Siu’s agent-driven workflow – built on OpenClaw, Instantly, and WhisperFlow – handles lead verification, sequence scoring, campaign distribution, and custom variable auditing end to end, with a human making the final judgment calls.
The Pulse:
- A single operator using OpenClaw’s agentic workflow replaces a 3-4 person cold email team, compressing a 2-3 week setup into one afternoon, according to Eric Siu of SingleGrain.
- Sequences are recursively scored against a simulated expert panel on persuasion and cold email mechanics, iterating until they clear a 90-out-of-100 threshold before human review.
- One campaign running on this infrastructure recorded a 16% positive reply rate – a 16x improvement over a baseline that Siu described as “pretty bad” in earlier campaign audits.
The friction at the center of modern cold email is not copywriting skill or lead quality. It is operational complexity: six or seven interdependent tasks that each require a different tool, a different specialist, and a different feedback loop. Eric Siu’s OpenClaw system addresses that complexity at the architecture level, not the task level – and the performance gap between the old approach and the agent-driven one is measurable in both time and reply rates.
The Operational Stack That Replaced a Full Team
The core mechanism is a single-brain agent system where OpenClaw acts as the orchestration layer, connecting to Instantly via API for campaign management and to Google Drive for sequence retrieval – so one operator can query, rewrite, score, and deploy campaigns without switching tools. Eric Siu describes the setup as an “agent fleet” operating around a “world brain,” where the human’s role shifts from executor to reviewer with taste and judgment.
The workflow begins with a performance audit. Siu queries the Instantly API directly through OpenClaw: “Tell me about the Instantly campaigns this week. What’s performing versus what’s not?” The agent surfaces reply rates, paused accounts, and historical benchmarks. The target threshold Siu uses is a 2-4% reply rate for cold outreach – anything below that triggers a sequence rewrite cycle.
When a rewrite is needed, the agent reads a Google Drive document containing successful historical sequences, then reformats and rewrites the new sequence to match the structural patterns of the high-performers. The agent does not guess at what works; it copies the architecture of sequences that have already demonstrated results, then adapts the copy to the new campaign context.
Parallel execution is a core feature of the architecture, not an add-on. While the agent was optimizing SingleGrain’s outbound sequences, Siu spun up a second instance for a separate campaign called “Picked Up Calls” targeting plumbing businesses. The two threads ran simultaneously, each receiving independent feedback and edits, without blocking each other.
The Real Takeaway: A single operator running parallel agent threads can manage two distinct campaign builds simultaneously, a task that previously required separate team members for each account.
Recursive Scoring and the 90-Plus Gate
Before any sequence reaches a human for review, the agent scores it recursively on a scale of 0 to 100 against a simulated expert panel calibrated to the specific skill domain – in this case, persuasion, marketing, and cold email mechanics – and only returns the output when it clears a score of 90 or above. This is not a single-pass quality check; it is an iterative inference loop that self-corrects until the threshold is met.
The expert panel is not a static rubric. It is constructed relative to what the campaign is selling and what technique it is using. A sequence for a plumbing business is scored differently than one for a B2B SaaS vendor. The mechanism forces the agent to apply domain-specific persuasion logic rather than generic copywriting heuristics.
Siu acknowledges the subjectivity: “Still a little subjective, but at least it uses a reported expert panel based on whatever it is that we’re offering.” The value is not perfection; it is that the agent filters out low-quality outputs before they consume human review time. By the time Siu reads a sequence, it has already passed a threshold that eliminates the worst 80% of drafts.
The practical output of this gate was a sequence for a plumbing client that opened with: “A plumber in Phoenix picked up 11 hours after calls last month that would have gone to voicemail, booked $4,200 in jobs from those calls alone.” Siu’s assessment: “That’s not bad. We’re looking for brevity when we write these emails. We hopefully want to have numbers in there, too.”
Why This Matters Now: Recursive scoring with a domain-specific expert panel means the human reviewer is evaluating polished candidates, not raw drafts, cutting review time by eliminating the bottom tier before it surfaces.
Custom Variable Auditing and the Automation Tell
One of the highest-use tasks the agent handles is custom variable verification – checking that every personalization token in a sequence resolves correctly against the actual lead list, because a fallback to a blank field is an immediate automation tell that kills deliverability trust. Siu caught this manually in one campaign and delegated the fix entirely to the agent.
The specific failure mode: if a sequence uses a first-name variable but the lead list has no first name entered, the email renders with a blank space where the name should be. “You already know immediately it’s automated,” Siu noted. The agent audited all custom variables across active campaigns, identified the mismatches, and applied corrections without Siu touching individual records.
A second variable issue surfaced in the SingleGrain sequences: a reference to “most B2B companies over 10 million” that Siu flagged as too specific and potentially limiting. He dictated the fix via WhisperFlow: remove the revenue threshold, keep the broader framing. The agent updated the sequence copy across all relevant campaign variants in one pass.
| The Conventional Approach | The Yacov Avrahamov Perspective |
|---|---|
| 3-4 team members each own one task (leads, copy, distribution, optimization) | One operator orchestrates all tasks through a single-brain agent system with parallel threads |
| Sequences are reviewed once by a human copywriter before going live | Sequences pass a recursive 0-100 scoring gate and must clear 90+ before reaching human review |
| Custom variable errors are caught post-send or by manual QA | Agent audits all variable tokens against the live lead list before campaign activation |
| Incentive offers are static (single Amazon gift card option) | Incentive architecture is A/B tested across three variants: Amazon, Visa, or charity donation |
| Campaign edits require logging into Instantly and updating fields manually | WhisperFlow converts spoken instructions into Instantly campaign changes in real time |
The Bottom Line: Custom variable auditing is not a cosmetic fix; a single blank-field render in a cold email destroys the personalization illusion and reduces reply rates across the entire send batch.
AI Content Generation Inside the Outbound Stack
The SingleGrain sequences that the agent produced reflect a specific AI content generation strategy: lead with the AI search visibility angle, name ChatGPT, Gemini, and Claude as the relevant engines, and open with a model built for Salesforce to establish credibility before broadening the claim. This is not generic outreach copy; it is authority building embedded directly in the cold email hook.
One sequence opened with: “Your buyers are asking ChatGPT and Perplexity for vendor recommendations before they ever contact you. Most marketers have no idea where their brand shows up in those results.” Siu later edited this to remove Perplexity and replace it with Gemini and Claude: “Those are the main ones. Perplexity isn’t really that relevant in my opinion.” The edit is a signal that sequence copy should reflect current AI engine market share, not a static list.
The revenue leak sequence for SingleGrain referenced a model originally built for Salesforce and Amazon to quantify AI search visibility gaps. The framing positions SingleGrain as a vendor with enterprise validation, then broadens the claim to all B2B companies. This is a textbook AEO strategy applied to cold email: anchor on a recognizable brand, then generalize the problem to the prospect’s context.
From an AI-powered SEO and content marketing automation standpoint, the approach mirrors what I see working in thought leadership content: specificity in the opening claim, a named enterprise proof point, and a clear cost framing (“revenue leak,” “money on the table”) that makes the prospect’s inaction feel expensive. The agent is not writing generic articles; it is producing expert articles calibrated to a specific ICP and a specific business outcome.
The Strategic Implication: Embedding ChatGPT citations and AI engine visibility framing into cold email copy is a direct application of GEO optimization logic to outbound, targeting prospects who already use AI search to evaluate vendors.
What the Human-in-the-Loop Actually Does
Siu is explicit that the agent system does not eliminate human judgment; it concentrates it – the operator’s role is to provide taste, catch ICP mismatches, and make calls the agent cannot make algorithmically, such as whether 166 leads loaded into a campaign is acceptable or whether agency-type companies fit SingleGrain’s target profile.
Two specific judgment calls appeared in the workflow. First, Siu noticed that one campaign had loaded only 166 leads instead of the expected volume, and that the leads appeared to be agencies rather than direct clients. He flagged both issues via voice dictation and instructed the agent to redistribute leads from other campaigns and verify that email accounts were not duplicated across the new V4 campaign set.
Second, Siu caught a sequence the agent produced that he had already tested and knew did not work. He told it directly: “I don’t like the two campaigns that you made over here. You’ve tested this, this doesn’t work.” The agent accepted the feedback and iterated. This is the correct human-in-the-loop model: not micromanaging every output, but intervening when institutional knowledge overrides the agent’s inference.
The entire session, including both the SingleGrain and Picked Up Calls campaign builds, took approximately one hour of Siu’s time. His framing: “This video took what, 10 minutes or so? Once I’m done with this, maybe to get this all set and done, maybe takes an hour of my day. But what about the other 7 hours?”
What This Means in Practice: The human-in-the-loop role in an agent-driven cold email system is not quality control on every output; it is pattern recognition and ICP judgment that the agent cannot replicate from historical data alone.
Frequently Asked Questions
How does OpenClaw connect to Instantly for campaign management?
OpenClaw hooks into Instantly via its API, allowing the agent to read campaign performance data, load sequences, set send limits, and redistribute leads without the operator logging into the Instantly dashboard. Siu’s workflow uses this connection to query reply rates, pause accounts, and push new sequences live, all from within the single-brain interface. The API integration means every campaign action is logged and reversible, not a manual override.
What is the “skill” mechanism and how does it prevent repeating instructions?
A “skill” in the OpenClaw system is a saved workflow or rule set that the agent can retrieve and apply to future tasks without the operator re-explaining the logic. Siu explicitly noted during the session that he wanted to save the gift card incentive unit economics rules as a skill so he would not have to repeat the same feedback in future campaign builds. This is analogous to a reusable prompt template, but stored at the agent orchestration layer rather than in a static prompt file.
How should email account distribution be managed to avoid inbox burnout?
Siu’s instruction to the agent was explicit: verify that email accounts assigned to each new V4 campaign are not duplicated across other active campaigns. Sending from the same account across multiple campaigns simultaneously accelerates domain reputation degradation and risks inbox placement. The agent was tasked with auditing account assignments and redistributing them so each campaign drew from a non-overlapping pool of warmed accounts. Siu also set the daily send limit to zero (unlimited) per campaign, relying on account-level warming caps rather than campaign-level throttles.
How does this approach compare to using OpenAI or Anthropic APIs directly for sequence generation?
Building directly on OpenAI’s GPT-4o or Anthropic’s Claude API gives you raw inference throughput and fine-tuning options, but requires you to build the orchestration layer, the Instantly integration, the Google Drive retrieval, and the scoring loop yourself. OpenClaw’s value is that the agentic workflow, the context window management across parallel threads, and the tool integrations are pre-built. For teams without an engineering function, the build-from-scratch approach on Azure OpenAI Service or AWS Bedrock would replicate the capability at higher latency and significantly higher development cost. The tradeoff is customization depth versus time-to-deployment.
Scale Your Authority With AI-Driven Content
AuthorityRank builds expert articles at scale that get cited by ChatGPT, Gemini, and Claude. One platform for AI content generation, SEO optimization, and thought leadership content that drives measurable authority.