Experts, personas, backgrounds — shaping WHO the agent is in a run

Three capability kinds shape an agent's identity and framing (as opposed to skills, which shape its methods). All three inject into the front of the system prompt, in a fixed order, before anything else the run adds:

## BACKGROUND: <id>        ← ambient truth, first, verbatim
<background body>

<persona text>             ← identity and voice

## EXPERT: <id>            ← labeled domain lenses
<expert body>

The order is load-bearing: backgrounds frame everything that follows, and putting them first keeps the prompt prefix stable across runs sharing a background (prompt caching). Experts are labeled sections precisely so the model can tell "domain lens" apart from "who I am".

Backgrounds — always-on ambient context

A background is a registry card whose body is standing context: company facts, product truth, an environment description. Injected FIRST and verbatim — no stripping — because ambient truth must arrive intact.

- type: harness
  goal: "Draft the Q3 partner update"
  mission: |
    Write the quarterly partner update.
  backgrounds: [acme-corp]

Use a background when every run in a family needs the same grounding ("what Acme is, our products, our tone"), instead of pasting boilerplate into each mission text. Backgrounds are per-owner resolvable: your private background card wins for your runs.

Persona — one voice per run

A persona is an identity/voice text (from the platform's persona store). Exactly one per run — a run speaks with one voice.

persona: analyst

Accepted spellings, all normalized to the same thing: persona: (canonical), digital_twin: (alias), and personas: [x] (plural from preset frontmatter — only the FIRST entry is used). If you list several personas, you did not get a blend; you got personas[0].

Personas also appear outside missions: plungeai_execute_agent/POST /v1/agents/{id}/execute accept a persona parameter, and CNL debate/validate blocks take persona/digital_twin per debater, judge, and validator — the same store, applied per role. That is the idiomatic way to run a multi-perspective panel: one debate block, different personas per debater.

Experts — labeled domain lenses (UI name: Specialists)

An expert is deep domain instruction material injected as a labeled ## EXPERT: <id> section — "reason like a securities lawyer", "apply SRE practices". Up to 3 load eagerly per run.

experts: [securities-law, python-pro]

Persona vs expert, the practical line: persona = who the agent IS (voice, identity — one). Expert = what the agent additionally KNOWS HOW to judge (lenses — up to three eager). A hedge-fund panel is personas in a debate; a compliance review is one persona plus a law expert.

Budgets and degradation (shared with skills)

All eager identity/instruction text — backgrounds + persona + experts + skills — shares one 24 KB budget, with per-kind caps: 3 backgrounds, 1 persona, 3 experts (5 skills). Over cap or over budget, items defer: the prompt lists them under ## AVAILABLE ON DEMAND and the agent can pull one mid-run with load_skill (matching type). Persona is never deferred — it either resolves or warns.

Missing ids never fail the run; they degrade to warnings:

• background "<id>" not found or empty
• persona "<id>" not found
• expert "<id>" not found

A run that "lost its voice" or ignored company context almost always has one of these warnings — check the run detail (plungeai-results-traces). The other classic cause: declaring more than the caps and assuming everything injected. Order lists by importance; the head injects, the tail defers.

Discover valid ids before declaring them — the degrade-to-warning behavior means a guessed id fails silently: plungeai_list_agents {kind: "experts", search: "…"} or {kind: "personas", search: "…"} (REST: GET /v1/discovery/search?kind=…&q=…).

Merge behavior in missions

Like every mission field, these merge per key with last-wins across pre-built card → workflow root → task, and arrays REPLACE rather than union (see plungeai-missions). So a task-level experts: [x] replaces the card's expert list — and an explicit empty experts: [] deliberately clears it. To ADD to a card's list you must restate the full list.

Authoring guidance

• Keep each body lean and self-contained: it lands in a prompt with everything else competing for 24 KB. A 15 KB background starves persona and experts.
• Backgrounds: facts, not instructions. Instructions belong in mission text or skills — a background that says "always do X" fights the mission's own framing.
• Experts: method and judgment criteria ("what a great X checks first"), not essays.
• Test identity injection cheaply: run a quick-effort mission whose goal is to introduce itself and state its operating context; the answer shows exactly which layers landed.