Agent Roster — Built and Proposed

The complete picture: 3 agents built, 7 proposed, 2 extensions to existing agents, and 2 roadmap items that should not be agents at all.

Back to the agents index · workflows

---

The test applied throughout

An agent is warranted where judgment on unstructured input is needed.

Where the work is deterministic — routing a message, buying media, counting events — an agent is a liability: slower, costlier, non-reproducible, and harder to audit. Several items the roadmap draws as agents fail this test, and are listed as integrations instead.

---

The nine

---

What exists today (1–3)

Three agents, all OpenAI-pinned, all tenant-scoped, all failing open.

1. Discovery Agent — ✅ built

Automatic, per item, on a background worker. Turns a raw scraped row into a browsable feed card: clean summary, category from the tenant allow-list, Indian state/UT, SEO keywords, authenticity and brand-fit scores.

Notably not all AI: authority scoring and persona matching are deterministic by design — authority is a property of the source, not the story, and an article never states its reader's age.

→ plain-English · technical

2. Compliance Agent — ✅ built

On demand, at two points, with different framing: judging a source article you might write from, and judging a draft about to publish under the brand's own name. Cites the tenant-authored rule it fired on, quotes the exact phrase, and blocks Send to CMS on high severity until a human acknowledges.

The only agent with a veto.

→ plain-English · technical

3. Creative Agent — ✅ built

Semi-automatic: a human queues a topic, generation runs in the background and stops at ready. Blogs, video scripts and hooks. Grades its own output and retries — quality repair, then up to two humanization passes — before a reviewer sees anything.

→ plain-English · technical

---

What to add (4–10), by tier

Tier 1 — buildable now, no new infrastructure

Tier 2 — one substantial dependency each

Tier 3 — blocked on the customer data layer

Building a segmentation agent before customer data lands would produce something with nothing to select from. The dependency is the blocker, not the agent count.

---

Two extensions — cheaper than new agents

Neither needs a new agent. Both are high value.

Script compliance — do this first

The Compliance Agent already takes subject: "source" | "draft". Scripts today get the deterministic keyword scan only, so a video script receives meaningfully less review than a blog post.

Adding subject: "script" is a prompt-framing change against proven plumbing — days, not weeks. It closes the widest review gap in the product for the least work of anything on this page.

Theme diversity — a Creative Agent capability

Page 4 of the roadmap asks for "Diff Approaches → Diff Themes". Today one prompt yields one angle; a second take means regenerating and hoping.

This is a generation parameter on an existing agent, not a new one.

Similarly, A4 Communication Creation on page 2 is the Creative Agent pointed at a segment — same judgment, different input. It does not need its own agent, only the segment to point at.

---

Two things that should not be agents

The roadmap draws these as agents (A5) or implies them. Both fail the judgment test:

Naming these explicitly matters: "agent" has become the default unit of work on the whiteboard, and applying it uniformly would put an LLM in the path of every outbound message.

---

Recommended sequence

NOW      Script compliance extension        days · closes the widest review gap
  ↓
1st      Localization Agent                 self-contained · no new integration
  ↓
2nd      Learning Agent                     highest leverage · makes 1-4 improve with use
  ↓
3rd      ORM Agent                          proven shape · vendor feed is the cost
  ↓
then     Visual · SEO Scale                 infra lift each
  ↓
blocked  Segmentation                       needs CRM first

Why Localization before Learning, despite Learning having more leverage: Localization needs nothing new and delivers on a strategy that is already half-built (geo tagging exists, micromarket drill-down is on three roadmap pages). Learning needs an event log that does not exist yet. Ship the unblocked one while that plumbing lands.

---

Scheduler is the odd one out. It is mostly deterministic and should stay that way — the calendar write, recurrence, timezones and blackout dates all need to be auditable row by row. Build the rules engine; add the ranker only when someone is visibly spending time on slot ordering.

Coverage after all ten

Even with all seven built, BTL and Program stay thin — because their blocker is a CRM and a channel layer, not a shortage of agents.