Build a multi-agent outreach system where specialised Claude agents prospect, write and send link pitches under one orchestrator — with a UK PECR and GDPR compliance layer built in.
A single prospecting agent finds and qualifies link targets. A multi-agent system runs the whole campaign — prospecting, writing and sending — as a coordinated crew of specialists, each doing one job well, all directed by an orchestrator. The performance gap between the two is not marginal. In Anthropic’s own internal evaluation, a multi-agent system built on an orchestrator and specialised subagents outperformed a single top-tier agent by over 90% on complex, breadth-heavy work, and platforms running multi-agent outreach crews have reported conversion lifts of up to sevenfold over one-dimensional single-agent setups.
This article is the build guide for that crew, applied to link outreach. It assumes you already understand single-agent prospecting and what backlinks are and why editorial relevance now outweighs raw authority. What it adds is the architecture for coordinating several agents, the handoffs between them, an honest cost model — multi-agent systems are expensive — and the part every generic guide ignores: a compliance layer built for the UK, where cold B2B outreach is governed by PECR and UK GDPR with fines reaching £17.5 million. Get the architecture right and you scale output; get the compliance layer wrong and you scale liability.
One principle frames everything below, and it comes from Anthropic’s own engineering guidance: find the simplest solution that works and only add complexity when it earns its place. A multi-agent system is complexity. You build it when one agent can no longer carry the campaign — not before.
The deliverable: the multi-agent outreach architecture
Here is the entire crew on one page. One orchestrator directs four specialist workers and routes everything through a human gate before a single email leaves the building. Read this and you have the system; the rest is detail.
| Agent | Role | Owns |
| Orchestrator | Campaign manager | Plans, dispatches work, integrates results, decides when done |
| Prospector | Finds + qualifies targets | Discovery, enrichment, relevance scoring, qualification |
| Writer | Drafts the pitch | Tailored opening email per qualified target, no sending |
| Compliance | Legal gatekeeper | PECR / UK GDPR check, suppression list, lawful-basis tag |
| Deliverability | Inbox protector | Sender hygiene, bounce checks, send scheduling, throttling |
| Human gate | Final authority | Reviews, edits, approves or kills — owns the send button |
Notice the shape: the orchestrator does no outreach work itself. It plans, dispatches and integrates — the project-manager role. The four workers are specialists with narrow mandates. The human gate is sovereign over the one irreversible action: sending. This division is the whole design, and it maps directly onto the broader playbook in our link building strategies guide.
Why a crew beats one big agent (and when it does not)
The instinct is to make one agent do everything. Resist it, for a specific mechanical reason. Anthropic’s multi-agent research system gains its edge mainly from parallel reasoning across multiple independent context windows — each subagent works in its own context, on its own slice, without drowning in the others’ detail. A single agent juggling prospecting data, copywriting and compliance rules in one context window degrades on all three. Specialists with clean contexts do not.
But multi-agent is not a free upgrade. Two honest caveats decide whether you should build one:
- It is expensive. Multi-agent systems consume roughly fifteen times the tokens of a standard chat interaction, so they only pay off when the value of the outcome clearly outweighs the spend. For outreach at any real volume, it does — but you must do the maths, which we do below.
- It suits breadth, not tight interdependence. Multi-agent excels at work that splits into parallel strands and is weaker on tightly coupled tasks. Outreach splits beautifully — prospecting, writing and compliance are largely independent — which is exactly why it is a good fit.
The market is moving this way regardless. Adoption surveys cited in the field put roughly one in ten organisations using AI agents today, over half planning to within a year and around 82% within three years, and Gartner forecasts cited by vendors expect B2B teams to cut prospecting and prep time by more than half through embedded generative AI. The operating model that holds up is consistent across every credible source: agents run the repetitive steps, humans own the exceptions and the relationships.
The orchestrator-worker pattern, explained for outreach
The architecture has a name and a proven shape. In the orchestrator-workers pattern, a central model dynamically breaks a task into subtasks, delegates them to worker models, and synthesises the results. Anthropic’s research product implements exactly this: a lead agent plans an approach and writes it to memory, then spins up several specialised subagents in parallel, each with its own context window and tools, and reconciles their findings — with a separate citation pass at the end to attribute every claim.
Two design decisions from that system carry straight over to outreach:
- Workers do not talk to each other. The orchestrator constrains the topology: every decision about what comes next lives in one place. Your Prospector never speaks to your Writer directly; the orchestrator passes a clean, structured handoff between them. This makes the system debuggable — when something goes wrong, there is one coordinator to inspect, not a tangle of peer-to-peer chatter.
- The isolation boundary is the key choice. Decide what each worker needs to know about the others’ work — and give it nothing more. The Writer needs the enriched prospect record and your offer; it does not need the Prospector’s rejected candidates or the Compliance agent’s internal rule table. Tight isolation keeps each context clean and each worker sharp.
In pseudo-code, the orchestrator loop is almost embarrassingly simple — which is the point:
| plan = orchestrator.plan(campaign_brief) # strategy -> memory qualified = prospector.run(plan.targets) # worker 1, parallel for prospect in qualified: verdict = compliance.check(prospect) # worker 2 (gate) if verdict != “lawful”: log_and_skip(prospect, verdict); continue draft = writer.draft(prospect, plan.offer) # worker 3 queued = deliverability.prepare(draft) # worker 4 review_queue.add(queued) # -> human gate orchestrator.integrate(review_queue) # synthesise + report |
The model routing matters here. Anthropic’s system pairs a powerful orchestrator with cheaper workers — it used a top-tier model as the lead and a mid-tier model for subagents. Do the same: Opus 4.8 to orchestrate and to settle hard judgement calls, Sonnet 4.6 for writing and tool-use, Haiku 4.5 for high-volume classification inside the Prospector and Compliance agents. You pay the premium only where it changes the outcome.
Single agent, workflow, or crew? Choosing the right complexity
Before building a crew, be honest about whether you need one. There are three tiers of sophistication, and most teams jump to the most complex when a simpler tier would serve. Match the tier to the problem:
| Tier | What it is | When it is enough |
| Single agent | One agent, one loop, many tools | Low volume; one campaign type; one person reviews everything |
| Workflow | Fixed code path orchestrating model calls | Predictable steps that never change order — a scored pipeline |
| Multi-agent crew | Orchestrator dynamically directing specialists | High volume; multiple concurrent campaigns; compliance + deliverability at scale |
The distinction that matters is between a workflow and a crew. Workflows orchestrate models through predefined code paths; agents dynamically direct their own process and tool usage. If your outreach steps are fixed and you can hard-code the order, a workflow is cheaper, faster and easier to debug than a crew — build that. Reach for a multi-agent crew only when the work genuinely cannot be predicted in advance: when the number and nature of subtasks depend on the input, when campaigns run concurrently across markets with different rules, or when no single context window can hold everything. Outreach at scale hits all three, which is why the crew earns its place — but a solo practitioner running one UK campaign a month almost certainly does not need one. Building complexity you do not need is the most common and most expensive mistake in this whole field.
The crew, agent by agent
The Orchestrator — campaign manager
The orchestrator receives the campaign brief (target topics, the offer, volume, market) and produces a plan it stores in memory. It then dispatches work to the specialists, monitors progress, handles exceptions the workers flag, and assembles the final report. Critically, it decides when the campaign has enough qualified, compliant, drafted pitches to stop. Memory persistence matters because a large campaign can blow past a single context window; Anthropic’s lead agent writes its plan down precisely so it does not lose the thread when token limits bite.
The Prospector — finds and qualifies
This is the single-agent prospecting build, slotted in as a worker: it discovers candidates, enriches them with metrics and contact data, scores them against a relevance-first rubric, and qualifies or rejects. It hands the orchestrator a clean list of qualified targets with structured records. Everything that makes a good standalone prospecting agent — relevance weighted over authority, a human-review tier for borderline cases — applies unchanged. If your campaign is guest contribution, its qualification aligns with the norms in our guide to guest posting for links.
The Writer — drafts the pitch
The Writer takes one enriched record plus the offer and produces a tailored opening email: a specific reference to the publisher’s recent work, a one-line statement of the asset on offer, and a single clear ask. It never sends. The discipline is to prompt against the generic template, because mass robotic outreach earns nothing and burns sender reputation. Constrain the Writer to personalise only from facts in the enrichment record — never from invention — so it cannot cite an article that does not exist.
The gap between a generic draft and a UK-tuned one is stark. Compare what an untuned Writer produces against a Writer briefed for the British market:
| GENERIC (untuned): Subject: Amazing data you’ll LOVE for your readers!! Hi there, I came across your awesome website and thought our incredible study would be a perfect fit. We’d love a backlink! Let me know! UK-TUNED: Subject: UK SaaS churn data — for your retention coverage Hi Sarah, your piece last week on B2B retention benchmarks was a useful read. We’ve just published original UK churn data by ARR band that may complement it — happy to share the figures if helpful. Source link below; no obligation. |
The first is a mass-campaign tell: superlatives, no name, no specific reference, an American register, and a transparent ask for a link. The second names the editor, cites a real recent piece, offers genuinely useful UK-specific data, and makes a low-pressure ask in British register. Only the second earns a reply — and only the second reflects the editorial norms we return to below. The Writer can produce the second at scale, but only if you brief it for restraint, locale and specificity, and only if the human gate confirms the recent-piece reference is real.
The Compliance agent — the legal gatekeeper
This is the agent the competition does not build, and in the UK it is the one that keeps you out of trouble. It checks every prospect against PECR and UK GDPR rules, screens against your suppression list, and tags each contact with a lawful basis before the Writer is allowed to draft. It is a hard gate: no lawful basis, no pitch. We give it a full section below because for a British outreach programme it is not optional. The same logic, adapted, governs outreach into the EU — see our guide to link building for European markets.
The Deliverability agent — inbox protector
Drafting a perfect pitch is wasted if it lands in spam. The Deliverability agent owns sender hygiene: it verifies the sending domain’s SPF, DKIM and DMARC records, runs bounce checks, and enforces opt-outs, then schedules and throttles sends so volume stays within human-plausible limits. It does not send on its own authority — it prepares a send-ready, scheduled package for the human gate to release. Think of it as the agent that protects the asset every other agent depends on: a clean inbox.
The Human gate — final authority
Every qualified, compliant, drafted, deliverable-checked pitch lands in a review queue. A human edits, approves or kills each one, then releases the approved batch to send. This is non-negotiable, and not for sentiment: AI-written outreach sent without human review is increasingly flagged by Google’s spam systems, and reputable teams keep editors for final control. The crew’s job is to make the human’s review fast and well-evidenced, not to remove the human from the loop.
Handoffs and shared state: how work moves between agents
Because workers never talk directly, the quality of the handoffs is the quality of the system. Anthropic’s subagents return condensed findings to the lead agent, often through a shared memory store rather than long conversational replies. Mirror that: each worker writes a compact, structured record to shared state, and the orchestrator reads it to decide the next step. Define the handoff schema explicitly so nothing is ambiguous:
| ProspectRecord = { domain, contact_name, contact_email, contact_type, primary_topic, relevance_score, authority, risk_flags, recent_article_title, recent_article_url, # for the Writer qualification: “qualify” | “review” | “reject” } ComplianceTag = { subscriber_class: “corporate” | “individual”, lawful_basis: “legitimate_interest” | “consent” | “none”, on_suppression_list: bool, verdict: “lawful” | “needs_consent” | “blocked” } DraftPackage = { subject, body, personalisation_source_fields, send_window, throttle_group, unsubscribe_link } |
These schemas are the contracts between agents. When a campaign misbehaves, you inspect the records passed at each handoff and find exactly which agent produced bad output — the same debuggability the orchestrator-worker topology buys you. Keep records condensed: pass the five fields the next agent needs, not the worker’s entire reasoning trace.
The UK compliance layer: PECR, UK GDPR and the suppression list
This is the section that separates a British outreach system from a generic one. Most multi-agent outreach content is written for a US context and ignores UK law entirely. In the UK, two regimes govern your link pitches: the Privacy and Electronic Communications Regulations (PECR) and UK GDPR, with the Information Commissioner’s Office (ICO) enforcing both and fines reaching £17.5 million or 4% of global turnover. The Compliance agent encodes the rules so they are applied to every single prospect, automatically, before anything is written.
| This is general information, not legal advice The rules below summarise widely-published guidance on UK B2B outreach. They are not a substitute for advice on your specific situation — confirm your approach with a qualified UK data-protection adviser before running a campaign at scale. |
Rule 1 — The corporate-subscriber exemption (PECR Regulation 22)
PECR Regulation 22 is the rule that makes most UK B2B link outreach lawful. Its corporate-subscriber exemption means you can send unsolicited B2B email to corporate subscribers — limited companies, LLPs, public bodies and Scottish partnerships — without prior consent, provided the message is relevant to the recipient’s role, identifiably from your organisation, and carries a clear, free opt-out. Most publishers and editorial outlets you pitch for links are corporate subscribers, which is why link outreach to UK media is generally on solid ground.
Rule 2 — UK GDPR still applies to every named person
Clearing PECR is not the end. UK GDPR still engages on every named individual recipient, so you need a lawful basis — usually legitimate interests — for processing their name and email. That means a documented legitimate-interests assessment and a privacy notice that tells the recipient who you are, why you are contacting them, how you got their data and how to opt out. The Compliance agent tags each prospect with its lawful basis so this is never an afterthought.
Rule 3 — Individual subscribers need consent
The exemption does not cover everyone. Sole traders, unincorporated partnerships outside Scotland, and personal-domain addresses are individual subscribers and require prior consent. A freelance journalist writing from a Gmail address is an individual subscriber; the limited-company magazine that commissions them is a corporate subscriber. The Compliance agent must classify the recipient correctly, because the rules flip entirely on that distinction.
Rule 4 — The suppression list is sacred
The most common PECR breach is mundane: re-emailing someone who already opted out. Checking every contact against a suppression list before sending is the single safeguard that prevents it. The Compliance agent screens every prospect against the suppression list as a hard first check — a suppressed contact is blocked before any other rule is even evaluated.
Rule 5 — Intent matters more than wording
You cannot dodge the rules with clever framing. The ICO position is that intent matters more than language — dressing a pitch up as a friendly note does not stop it being marketing if it promotes your business. A link pitch that exists to win you a backlink is marketing, full stop. Build for that reality rather than against it.
Encoded as a decision table, the Compliance agent’s core logic is simple and auditable:
| Recipient | Subscriber class | Verdict (B2B link pitch) |
| Named editor @ limited company / LLP / public body | Corporate | Lawful — legitimate interest + opt-out |
| Generic alias (editor@, info@) at a company | Corporate | Generally lawful — treat with extra care |
| Sole trader / partnership (exc. Scotland) | Individual | Needs prior consent |
| Personal-domain address (Gmail, Outlook) | Individual | Needs prior consent |
| Anyone on your suppression list | Either | Blocked — do not contact |
In pseudo-code, the gate runs suppression first, then classification, then lawful-basis assignment:
| def compliance_check(prospect): if prospect.email in suppression_list: return “blocked” cls = classify_subscriber(prospect) # corporate | individual if cls == “individual” and not has_consent(prospect): return “needs_consent” # route out of campaign record_legitimate_interest(prospect) # log the basis return “lawful” |
Two further habits keep you clean: never buy lists — broker-collected consent does not transfer to you, and purchased lists are high-risk — and honour every opt-out promptly, adding it to suppression the moment it arrives. ICO enforcement focuses on the egregious: high-volume spam, missing unsubscribe links, ignored opt-outs. Stay off that list and you stay off the ICO’s radar.
The cost model: what a multi-agent campaign actually costs
Style demands the numbers, and honesty demands we start with the bad news: multi-agent is roughly fifteen times the token cost of a single chat turn for an equivalent unit of work. The good news is that the absolute numbers are still trivial against human time, because you route the expensive model narrowly and batch the rest. Take a campaign of 200 qualified prospects taken all the way to send-ready, drafted and compliance-checked. Using Anthropic’s published per-token rates (quoted in pounds for convenience) and the Batch API’s 50% discount on asynchronous work:
| Component | Model | Volume | Cost |
| Pitch drafting | Sonnet 4.6 | 200 drafts | ≈ £1.20 |
| Compliance checks | Haiku 4.5 | 200 checks | ≈ £0.25 |
| Orchestration overhead | Opus 4.8 | campaign-wide | ≈ £3–5 |
| Total per campaign | — | 200 pitches | ≈ £5–7 |
Five to seven pounds to take 200 prospects to send-ready, with prospecting costs (covered separately) adding only a little more. Figures are illustrative and your token mix will differ, but the order of magnitude is the message: the 15x multi-agent premium is real and still leaves you paying single-digit pounds for work that previously consumed a two-person team for a week. The levers that keep it there are the same three every time — batch the asynchronous passes, cache the fixed instructions and rubrics, and route the premium model narrowly. For the wider tooling economics, see our comparison of link building tools.
| Throughput is the real prize The cost saving is nice; the throughput change is transformational. A multi-agent crew can prospect, write and compliance-check hundreds of pitches in the time a human team triages dozens. The human hour shifts entirely to the gate — reviewing well-researched, pre-checked drafts — which is the highest-value place for human judgement to sit. You are not replacing the team; you are removing everything that did not need them. |
A worked campaign: one UK link push, end to end
It helps to watch the crew run once. Here is a single UK campaign from brief to review queue, with the decision each agent makes.
- The brief. You hand the orchestrator a campaign: win editorial links from UK B2B SaaS and martech publications for an original UK SaaS-churn data study. Target volume, 200 qualified pitches. The orchestrator drafts a plan — topics, competitor sources, the offer, the market rules — and writes it to memory.
- Prospecting. The Prospector runs parallel discovery across query expansion, competitor backlink mining and AI-citation checks, enriches a pool of several hundred UK candidates, and scores them relevance-first. It returns roughly 240 qualified records — a deliberate over-supply, because compliance will cut some.
- Compliance. The Compliance agent screens all 240. It blocks 12 on the suppression list, routes 18 out as individual subscribers — freelance journalists on personal-domain addresses needing consent — and tags the remaining 210 as lawful under legitimate interest, each with the basis logged. The 18 are not discarded; they are flagged for a consent-based approach later.
- Writing. For the 210 lawful targets, the Writer drafts a tailored pitch each, referencing a specific recent article from the publication and offering the churn study. Personalisation draws only from enrichment facts, so nothing is invented.
- Deliverability. The Deliverability agent confirms the sending domain’s authentication is healthy, verifies addresses, removes a handful that bounce-test as risky, and schedules the survivors across a throttled, human-plausible send window with unsubscribe links in place.
- The gate. Around 200 send-ready, compliant, personalised pitches land in the review queue with their evidence attached. A human spends a focused session approving, sharpening and killing — reviewing 200 pre-checked drafts in the time it once took to research and write a dozen.
The whole run costs single-digit pounds of compute and a few hours of human judgement, against what was previously a week of two people’s time. Every irreversible decision — what gets sent, to whom, in what words — still belongs to a person. The crew removed the grind, not the judgement.
Beyond the law: UK editorial norms that win links
Compliance keeps you legal; it does not win links. Winning UK editorial links also means writing for British editorial culture, which differs from the US-centric defaults most AI outreach tools assume. Encode these norms into the Writer’s brief and your reply rates climb.
- Restraint over hype. UK editors are allergic to the breathless, superlative-laden American pitch. A measured, specific opening that respects their time outperforms enthusiasm. Prompt the Writer for understatement, not exclamation marks.
- UK spelling and idiom. “Optimise”, not “optimize”; “whilst” reads naturally; £ not $. A pitch in American English signals a mass, untargeted campaign to a British desk — the opposite of what you want. Set the Writer’s locale explicitly.
- Genuine localisation of the asset. UK-specific data — figures in pounds, regional breakdowns, British market context — is far scarcer in editorial inboxes than US-default data, and far more linkable for a UK publication. The Prospector should privilege UK-framed angles and the Writer should lead with them.
- The dense regional and trade press. Britain has an unusually deep ecosystem of trade journals, sector blogs and regional outlets that move fast and credit inconsistently. They are lower-friction targets than the nationals and a strong proving ground for a new crew.
- Relationships compound. UK trade editors are a small, well-networked world. A clean, useful first pitch that respects the rules builds a relationship that pays off across future campaigns — which is exactly why the human gate and genuine value matter more than volume.
This cultural layer is the difference between a technically-compliant campaign that gets ignored and one that earns coverage. It sits on top of the strategic fundamentals in our link building strategies guide, applied with a specifically British accent.
Failure modes: what breaks in a multi-agent system
Coordinating agents introduces failure modes a single agent never has. Anthropic’s own team hit most of these in development; engineer against them from the start.
- The orchestrator over-spawns. Early versions of Anthropic’s system spawned dozens of subagents for simple queries and chased sources that did not exist. Give the orchestrator explicit limits and clear stop criteria, or it will burn tokens coordinating work that did not need splitting.
- Vague delegation. Each worker needs a clear objective, an output format, and task boundaries. An under-specified handoff produces a worker that guesses — and guessing compounds across a crew.
- Hallucinated personalisation. The most damaging outreach failure: a draft citing an article the publisher never wrote. Constrain the Writer to enrichment facts only, and let the human gate catch the rest. This alone justifies the gate.
- Compliance bypass under load. If the orchestrator can route around the Compliance agent when it is busy or errors, it eventually will. Make the compliance gate a hard dependency: no lawful tag, no draft, no exceptions.
- Context bloat in the orchestrator. A campaign-long orchestrator conversation balloons fast. Persist the plan to memory, pass condensed records between agents, and process prospects in independent batches rather than one ever-growing thread.
- Autosend creep. The temptation to let the Deliverability agent send directly grows with confidence. Never give in. The human gate owns the send button, permanently.
Measuring the crew: per-agent and system KPIs
A multi-agent system needs both per-agent metrics (to find which specialist is underperforming) and system metrics (to judge the campaign). Track these and review weekly while tuning:
| Metric | Agent / system | What it reveals |
| Qualification rate | Prospector | Discovery breadth vs rubric tightness |
| Compliance-pass rate | Compliance | Health of your list sources — low pass means bad inputs |
| Reply rate on drafts | Writer | Whether the pitches land with editors |
| Bounce + spam rate | Deliverability | Sender hygiene — rising numbers are an early warning |
| Human-override rate | Gate | How often reviewers reverse the crew — your accuracy gauge |
| Placement rate per sent pitch | System | The bottom line: links won ÷ pitches sent |
| Cost per won link | System | Total compute + tools + human hours ÷ links |
The compliance-pass rate is the one teams overlook and should not: a low pass rate is not a compliance problem, it is a sourcing problem telling you your prospect inputs are full of individual subscribers or suppressed contacts. Fix the input and the rate recovers. For benchmarks on healthy outreach acceptance and link velocity, cross-reference our 2026 link building statistics.
Questions teams ask before building a crew
Is a multi-agent system overkill for my campaigns?
Probably, if you run one campaign at a time and one person reviews everything — a single agent or a fixed workflow will serve you better and cost less. The crew earns its complexity once you are running concurrent campaigns at volume, or once compliance and deliverability at scale exceed what one agent and one person can hold. Start at the lowest tier that works and graduate only when you feel the ceiling.
Does the UK compliance layer slow everything down?
It adds a gate, not a delay you notice — the Compliance agent checks each prospect in well under a penny of compute, in parallel with the rest of the pipeline. What it removes is the far slower, far more expensive scenario of an ICO complaint or a damaged sending domain. A compliance pass that blocks a fifth of your list is the system working, not failing.
Can the crew send emails on its own?
It can, technically, and you should not let it. The whole field converges on the same model: agents run the repetitive work and humans own the exceptions and the relationships. The send is the one irreversible, reputation-bearing, legally-significant action in the whole pipeline. A human owns it, always.
Which models should each agent run on?
Opus 4.8 for the orchestrator and for genuinely hard judgement; Sonnet 4.6 for the Writer and any tool-heavy worker; Haiku 4.5 for high-volume classification inside the Prospector and Compliance agents. The premium model should touch the minority of work where its judgement changes the outcome — everything routine runs cheap.
How is this different from an off-the-shelf AI SDR tool?
Packaged tools give you a fixed crew and someone else’s logic, usually built for a US sales context. Building your own means the compliance layer is genuinely UK-native, the rubric encodes your definition of a good link, and the costs are raw compute rather than per-seat licences. For repeatable, standardised campaigns a packaged tool may be enough; for a programme where UK-specific judgement is your edge, owning the crew is worth the build.
Your Monday-morning build path
You do not build the full crew on day one — that violates the simplicity principle. Build it in layers, proving each before adding the next. Start with the simplest thing that works and add agents only when one can no longer carry the load:
- Week 1 — Orchestrator + Prospector + Writer, with a human doing the compliance check and the send by hand. This is a working system that already saves hours; prove the qualification and reply rates first.
- Week 2 — Add the Compliance agent. Encode the suppression check, subscriber classification and lawful-basis tagging. This is the highest-value addition for any UK programme and the one that lets you scale safely.
- Week 3 — Add the Deliverability agent. Wire up SPF / DKIM / DMARC verification, bounce checks and send throttling so volume stays inbox-safe.
- Week 4 — Harden the handoffs. Lock the record schemas, add the orchestrator’s spawn limits and stop criteria, and instrument the per-agent KPIs.
- Ongoing — Tune from the human-override and compliance-pass rates. Every overridden verdict is a labelled example of where an agent is wrong; feed batches of them back into the relevant agent’s instructions.
Build it in that order and you end up with a system that scales output without scaling risk — a crew that prospects, writes and prepares hundreds of compliant, personalised pitches while a human owns every decision that matters and the single irreversible action of sending. That is the whole promise of multi-agent outreach: not removing the human, but removing everything that was never a good use of one. From here the cluster goes deeper still — retrieval-augmented personalisation that makes the Writer sharper, and vector-based prospect matching that makes the Prospector faster — but they all plug into this same orchestrated crew. Get the architecture and the compliance spine right first, and the rest is upgrades.
