Most teams shipping their first AI agent in 2026 expect the second to be easier. It usually is not, and the reason is rarely the agent itself. It is the substrate beneath the agent, and the operator muscle around the substrate, and the audit trail that needs to capture everything both agents do without anyone editing the schema as the system grows.
This is the playbook for going from one agent to five, derived from Legacy’s 12-month progression on Atlas (email-only deflection at go-live, then web chat, then voice agents, then integrated agents, ending year one at 8x total case volume on one substrate). It generalizes to any team starting on a substrate-first platform.
The principle: the substrate compounds; the agents do not
The first agent is mostly about the substrate. Curating the knowledge base. Configuring the approval seat. Standing up the audit log. Wiring the first connector. Teaching the operator what the system does well and what it gets wrong. The agent itself is the smallest thing in the deployment.
The fifth agent is the opposite. The substrate is mature. The operator trusts the system. The audit log has 12 months of evidence. The fifth agent reuses the knowledge base, the source-receipt machinery, the approval seat, the audit trail, and the connector catalogue that the previous four agents built up. The fifth agent ships in days; the first agent took weeks.
This is the asymmetry. If the substrate compounds, the fifth agent is faster than the first. If the substrate does not (because each agent is its own platform), the fifth agent is harder than the first, and most teams stall at the second.
Agent #1: the smallest possible thing
Pick the channel that is hardest to embarrass yourself on. For Legacy, that was email-only support deflection. Async, drafts reviewed before send, operator catches everything, no real-time pressure.
The work in months 1-3:
- Curate the first version of the knowledge base. Source documents, document versions, approval boundaries.
- Configure the approval seat. Who reviews, what they see, what they edit, what they approve.
- Stand up the audit log. Schema is fixed at the substrate level; the team just confirms the right tenant scoping and retention.
- Wire the first connector. For email-only deflection, this is just the email connector and the knowledge-base connector.
- Train the operator on what the system does well and what it gets wrong. Two weeks of close observation; iterate the knowledge base based on what the operator catches.
At the end of agent #1, the team has substrate. Substrate is what compounds. The agent is a thin layer on top.
Agent #2: the second channel
The most common second agent is the same workflow on a different channel. Legacy added web chat (the channel where visitors ask questions in the moment instead of writing email). The knowledge base transferred directly. The source-receipt machinery transferred directly. The approval seat behavior transferred directly with a small adjustment for real-time interaction.
The work in months 4-6:
- Add the chat-widget channel.
- Adjust the approval seat for real-time review (the operator now confirms patterns rather than reviewing every individual draft pre-send).
- Expand the knowledge base based on the new pattern of questions visitors ask.
- Iterate on which questions auto-route to chat versus get escalated to a human.
The risk on agent #2: assuming it is a copy of agent #1. It is not, because the channel is different and the operator’s daily work is different. The reusable part is the substrate.
Agent #3: a new channel that is harder than chat
Voice. SMS. Mobile push. The pattern: a channel where the agent is the first responder and the operator review is a flag-and-iterate pattern, not a draft-and-approve pattern.
Legacy added voice agents on inbound calls. The work in months 7-9:
- Add the voice connector. Speech-to-text on inbound, text-to-speech on outbound, real-time grounding against the knowledge base.
- Shift the operator’s review pattern: from approving each draft to flagging recordings where the agent said something that should be added to the knowledge base’s “do not say” list.
- Expand the audit log to capture the voice transcripts and the model decisions made in real time.
The substrate change is small (one new connector, one new review pattern). The capability change is large (the team can now answer phone inquiries 24 hours a day).
Agents #4 and #5: integrated agents
The big lift. Agents that read from and write to the systems the team already uses. Legacy added agents integrated with their CRM and their kit-ordering system, both built from connectors in the Clarm catalogue.
This is the step most teams stall on if their substrate is not mature. Integrated agents can break things in production systems. The reason it works at this stage of the progression: by month ten, the team has spent nine months trusting the approval queue, trusting the source receipts, trusting the audit log. The substrate has earned the right to write to production systems.
The work in months 10-12:
- Enable the connector for the integrated system. Configure write-side approval (no agent writes anything to the system without an operator-approved draft).
- Define the workflow templates. What does the agent do, in what order, with what data, requiring what approvals.
- Run the agent in shadow mode for two weeks. Drafts go to the approval seat but nothing actually writes to the production system. Operator confirms the drafts are right.
- Flip to live. Drafts approved by the operator write to the production system. Audit log captures every action.
At the end of month 12, the team has five agents in production across four channels (email, chat, voice, integrated) on one substrate. The fifth agent took two weeks; the first agent took twelve.
What compounds at each stage
- Knowledge base. Each agent adds new approved documents, new clarifications, new boundary cases. The knowledge base is more valuable on agent #5 than on agent #1.
- Operator muscle. The team that has spent 12 months in the approval seat reads drafts faster, catches issues faster, and identifies patterns faster than the team on day one.
- Audit history. A year of audit evidence is the document compliance teams trust. The audit log on agent #1 is empty; on agent #5 it is the evidence package the regulator already trusts.
- Connector catalogue (specific to the deployment). Each connector the team enables is one more system future agents can use without engineering work.
- Workflow templates. Each agent encodes a workflow procedure that future agents can adapt or extend.
What does not work
- Shipping all five agents in parallel. Substrate is not yet trusted. Operator muscle is not built. Audit history is empty. Most parallel five-agent launches end up with four shadow deployments and one production deployment that is doing the work of the original demo.
- Building each agent on a different platform. The substrate does not compound; the operator has to learn five different approval patterns; the audit log is fragmented; the connector work is duplicated.
- Treating the first agent as a throwaway proof-of-concept. The substrate work in agent #1 is the work that makes agents #2 through #5 possible. Throwing it away means starting over on substrate while the board thinks you are starting agent #2.
The decision that decides the whole progression
The single decision that decides whether the first-to-fifth progression is faster on agent #5 than agent #1 is the substrate. Substrate-first platforms (Atlas, by design) compound. Feature-first platforms (most of the 2025 cohort) do not, because each new agent reopens the governance, the audit, the approval, the source-receipt, and the tenant-isolation conversations.
For the substrate definition, read What Is Atlas?. For the specific 12-month case study this playbook generalizes from, read How Legacy Went 8x in 12 Months on Atlas. For the connector catalogue that makes agents #4 and #5 easy, read The Clarm Connector Catalogue. For the buying-side framing, read The Agent-Deployment Buying Guide.