Chapter 9

Multiple Agents and Subagents

Your research specialist and SEO manager from Chapter 8 now need teammates. A SERP analyst to examine top-ranking pages. A writer to draft the article. A reviewer to catch gaps before anything publishes. And these agents must work in the right order, on the right files, without overwriting each other's output.

This chapter answers three questions: when should you create a new profile versus spawning a temporary subagent? How do you coordinate multiple agents so they do not step on each other? And how does this work in practice for your SEO workflow?

Two kinds of agents

versus

Use a profile when

  • The role repeats across sessions and should accumulate expertise — keyword research, content briefs, SERP analysis
  • The agent needs its own memory, skills, and toolset configuration

Use a subagent when

  • The task is a one-off — run a single search, verify a single fact
  • You want parallel execution or fresh context with no contamination from the parent's conversation

Profiles (Ch3) for durable roles, subagents for one-off tasks. Profiles compound: they get sharper over time because they carry forward memory, skills, and corrections. Subagents are disposable: they do clean work because they start with no baggage, but they also start with no accumulated knowledge. The parent profile spawns a subagent via delegate_task, the subagent runs in its own context window, returns a summary, and disappears. Subagents can run in parallel but are not durable — if the parent is interrupted, the child is cancelled.

Coordination

: a shared board for multiple workers

Two agents can hand off directly — the research specialist saves a file, the SEO manager reads it. At three or more, you need a coordination system. Kanban provides a shared task board backed by SQLite, independent of any single agent process. The flow: create a task → the dispatcher assigns it → the agent claims and works (sending heartbeats) → the task is marked complete or blocked → the next linked task unlocks. If heartbeats stop, the dispatcher reclaims the task for another agent. No manual intervention needed.

bashVerified
$ hermes kanban list

  #1  keyword-research    [complete]  research-specialist
  #2  serp-analysis       [complete]  serp-analyst
  #3  content-writing     [complete]  content-writer
  #4  draft-review        [in_progress]  seo-manager

$ hermes kanban show 4

  Task #4: Draft Review
  Assignee: seo-manager
  Status: in_progress
  Heartbeat: 12s ago

Kanban prevents file conflicts in three ways. First, atomic task claiming — only one agent can claim a task at a time, so two researchers get separate tasks instead of the same one. Second, sequential handoff by design — the next task does not become ready until the previous one completes, enforcing the right order without manual oversight. Third, profile isolation — each profile has its own working context, and the board ensures they are not working on the same task. The board coordinates access; the profiles stay in their lanes.

File naming conventions add a safety layer: the research specialist writes to keywords.md, the SERP analyst to serp-analysis.md, the writer to draft.md. Separate files for separate roles makes conflicts structurally impossible.

Your SEO workflow

From two agents to a coordinated team

Each of these four roles repeats weekly and accumulates expertise — all are profiles, not subagents. The SEO manager shifts from strategist-drafter to pure orchestrator and reviewer, holding your judgment at the top of the chain.

Research specialist

Finds keywords using web search. Writes to keywords.md.

Toolsets: Web search, file read/write

SERP analyst

Reads keywords.md, analyzes top-ranking pages, reports gaps and formats. Writes to serp-analysis.md.

Toolsets: Web search, file read/write

Content writer

Reads keywords + SERP analysis, drafts article. Writes to draft.md — does not publish.

Toolsets: File read/write, web search

SEO manager

Reviews drafts against your style rules. Manages the kanban board. Nothing publishes without sign-off.

Toolsets: File read/write, kanban

Split roles prevent context pollution. If one agent both researches keywords and drafts articles, the search results fill its context window — leaving less room for the writing task. One profile, one job, one output contract.

When to use subagents

Subagents for parallel one-off work

The four profiles handle the recurring weekly cycle. But sometimes you need one-off work — a parallel search for two topic clusters, a quick verification that a published article is still ranking. The research specialist spawns two subagents in parallel, each searching a different topic, then merges the results. Subagents are synchronous and not part of the kanban flow — they are extra hands for a single job, gone when the job is done.

Not sure whether to use a subagent or a kanban task? Three questions:

  • Will this task repeat? Yes → profile on the kanban board. No → subagent is fine.
  • Does it need its own toolset or memory? Yes → profile. Subagents inherit the parent's toolset.
  • Must it survive interruptions? Yes → kanban task (durable, reclaimed if agent crashes). No → subagent (synchronous, cancelled if parent interrupted).

Most SEO workflows end up with a mix: profiles on the board for the weekly cycle, subagents for one-off research and parallel fact checks. The mix gives you both compounding expertise and flexible throughput.

Practical safeguards

Keeping agents from stepping on each other

Kanban coordinates task access. File conventions coordinate data access. Three rules make conflicts impossible by design:

Rule 1 — Separate output files per role

Each profile writes to its own file. The SEO manager reads all files but writes only to review.md. No two profiles write to the same file.

Rule 2 — Read-only for upstream data

The SERP analyst reads keywords.md but never modifies it. The content writer reads both upstream files but never modifies either. Downstream agents treat upstream outputs as read-only inputs.

Rule 3 — Kanban gates the read

A profile does not read the upstream file until the board says the upstream task is complete. Sequential access through task status, not file locks.

Takeaway

Designing for scale without losing control

Expanding from two agents to four changes the problem from "how do I make each agent sharp?" to "how do I make the team work together?" Profiles with skills and output contracts make each agent sharp at its own job. Kanban makes the team work together without collisions. Subagents add flexible throughput for one-off work. And the SEO manager — holding your judgment — sits at the top of the chain, reviewing every draft before it goes anywhere.

You need to add a weekly performance review to your SEO team: an agent that checks whether your published articles are ranking for their target keywords and reports any drops. Should this be a new profile on the kanban board, or a subagent spawned by the SEO manager? Why?