Editorial Answers Make Declarative Agents Predictable

Some agent answers should not be improvised. That sounds obvious until a user asks the same high-stakes question three different ways and your declarative agent gives three technically plausible answers with three slightly different meanings. Editorial answers, shipped in the 1.7 declarative agent manifest schema, fix that by letting builders reserve exact responses for known questions while still letting semantic matching catch natural language variations.

What Editorial Answers Are Trying to Fix

A declarative agent is wonderful at synthesizing. It reads instructions, uses capabilities, calls actions, grounds in knowledge sources, and produces a helpful answer. That is the magic.

It is also the problem when the answer must be boringly consistent.

Think about an HR onboarding agent. A new hire asks:

That answer is reasonable. It is also not the answer the organization may want. Maybe the policy is stricter. Maybe the right response must point to a specific procurement process. Maybe Legal reviewed the wording and said, “Please, for the love of compliance, use exactly this paragraph.”

Editorial answers are designed for that kind of known, repeatable answer. The idea is simple: define question and answer pairs, then let the platform select the right answer when the user’s question is semantically close enough.

Editorial Answers Are Not Better Instructions

Instructions shape behavior. Editorial answers shape a specific response.

That distinction matters. If you need the agent to behave with a certain tone, stay in scope, cite sources, or follow a workflow, use effective agent instructions. If you need the agent to return a precise answer for a known question, editorial answers become the more interesting control surface.

A good mental model:

Control surface	Use it for	Do not use it for
Instructions	Role, tone, scope, workflows, formatting	Exact policy paragraphs that must never drift
Knowledge sources	Grounding in documents, files, mail, Teams, connectors	Forcing one approved answer shape
Evals	Measuring whether behavior improved	Delivering the answer at runtime
Editorial answers	Known questions with approved responses	Broad reasoning or open-ended synthesis

The reason this is exciting is not that editorial answers make agents sound nicer. It is that they make specific answers more reliable.

Editorial Answers Land in v1.7

The 1.7 declarative agent manifest schema, published on 2026-05-14, adds editorial_answers as a top-level optional property. Set version: "v1.7" in your declarativeAgent.json and add the object. That is the whole onboarding ceremony.

{
  "$schema": "https://developer.microsoft.com/json-schemas/copilot/declarative-agent/v1.7/schema.json",
  "version": "v1.7",
  "name": "HR Onboarding Buddy",
  "description": "Helps new hires get answers about onboarding policies.",
  "instructions": "You are an HR onboarding assistant who answers questions about company policies for new hires.",
  "editorial_answers": {
    "answers": [
      {
        "question": "Can contractors expense home office equipment?",
        "answer": "Contractors are not eligible for home office equipment reimbursement through the employee expense process. Please contact your vendor manager for contract-specific purchasing guidance.",
        "similarity_thresholds": {
          "min": 3,
          "max": 8
        }
      }
    ]
  }
}

Three rules worth memorizing before you write a hundred of these:

• Either answers or url, not both. The editorial_answers object accepts an inline answers array (shown above) or a url pointing at an external document that contains the same pairs. Pick one per manifest.
• 300 answers maximum. The answers array caps at 300 entries. That cap is a feature, not a limitation. Editorial answers are for the questions that absolutely must not drift, not a replacement for your knowledge sources.
• similarity_thresholds is optional. When you do set it, min and max are integers from 0 to 10 inclusive. Think of min as “close enough to consider” and max as “obviously the same question.” Omit the property and the platform uses its defaults. Only tune the answers that misfire.

Validate With Evals

The schema lets you ship the feature. Evals let you trust it. Treat every editorial answer as a test case so you can prove the answer fires for the questions it should and stays quiet for the ones it should not:

{
  "schemaVersion": "1.2.0",
  "default_evaluators": {
    "Similarity": {},
    "Relevance": {}
  },
  "items": [
    {
      "name": "Contractor equipment reimbursement answer",
      "prompt": "Can contractors expense home office equipment?",
      "expected_response": "Contractors are not eligible for home office equipment reimbursement through the employee expense process. They should contact their vendor manager for contract-specific purchasing guidance."
    },
    {
      "name": "Semantic variation of the same policy question",
      "prompt": "I am a contractor. Can I submit a monitor and chair as expenses?",
      "expected_response": "Contractors are not eligible for home office equipment reimbursement through the employee expense process. They should contact their vendor manager for contract-specific purchasing guidance."
    }
  ]
}

The evals become your safety rail: the editorial answer should improve similarity on its intended questions without making the agent worse on adjacent ones.

Where Editorial Answers Fit

Editorial answers should be narrow. If you need hundreds of them to make the agent useful, you probably have a content architecture problem, not an agent configuration problem.

Use them for:

• Regulated answers: HR, legal, finance, compliance, and support responses where wording matters.
• High-volume known questions: The same question arrives constantly, with small wording changes.
• Launch-critical answers: Questions executives, sellers, partners, or customers will definitely ask during rollout.
• Policy boundary answers: The agent must say “no” clearly, consistently, and without improvising.

Avoid them for:

• Open-ended analysis.
• Questions that require fresh retrieval from documents.
• Answers that change daily.
• Anything better handled by an API plugin, MCP plugin, or knowledge source.

This is the same lesson as agent evals: quality improves when you stop treating agent behavior as vibes and start treating it as an asset you can inspect.

A Before and After Worth Wanting

The goal is not to make the agent robotic. The goal is to make the right answer show up when the user is clearly asking a known question.

Before editorial answers, a known policy question might get a synthesized answer:

After editorial answers, the same intent returns the approved wording:

That is the value: not fancier prose, just less drift.

The Value: Specific Answers Without Guesswork

• Predictable policy responses: Known questions can map to reviewed wording instead of fresh synthesis every time.
• Cleaner quality reviews: Evals can compare answer behavior before and after the editorial answer is introduced.
• Better separation of concerns: Instructions guide the agent, knowledge grounds the agent, editorial answers reserve exact responses.
• Safer launches: Teams can identify the questions that must be boringly correct before a pilot goes wide.

Editorial answers are worth reaching for because they fill a gap between instructions and evals. Ship the answers that absolutely must not drift, let evals keep them honest, and leave the rest to the agent’s normal synthesis.

Resources

• Declarative agents overview
• Declarative agent manifest schema
• Declarative agent manifest v1.7
• v1.7 JSON schema
• Writing effective agent instructions

---

The Manifest is where we turn agent configuration into something you can reason about, review, and ship without crossing your fingers.