Schema Mind
How the agent's output contract shapes its thinking. Every field in BriefResult forces a decision. Vague fields produce vague output. Clear fields produce clear thinking.
Design the output shape before the prompt
Before the agent sees a single article, someone decided the shape of its output. That shape — the schema — is the hidden architecture of agent cognition. The agent doesn't think freely. It thinks into a form. Every field is a question it must answer. Every type is a constraint. Leave a field out and the agent never considers that dimension.
The 10 Fields of BriefResult
Each field demands something specific from the agent. Fields scroll-reveal as you read.
article_type
enum → string
Forces: categorization, not just description.
The agent must decide whether this is an industry_shift, a product_launch, a research_paper, or something else. This shapes recall: different types connect to different past briefs.
core_thesis
required → string
Forces: distillation to one claim.
The agent cannot summarize. It must find the single argument the article makes. If the article makes no argument, this field exposes that.
key_insight
required → string
Forces: finding what's actually new.
Different from the thesis. The thesis is what the article argues. The insight is what the agent noticed that isn't obvious. This field trains the agent to look past the surface.
reading_depth
enum: skim | deep
Forces: honesty about effort invested.
The agent must admit when it skimmed. This gates downstream decisions: a skimmed article shouldn't be used to support strong claims.
reading_depth_reason
required → string
Forces: justification for the depth choice.
The agent cannot just say "skim." It must explain why. This creates an audit trail: was the skim justified or lazy?
familiarity_level
enum: novel | familiar | well_worn
Forces: calibration against existing knowledge.
A novel article gets different treatment than a familiar one. This field prevents the agent from treating everything as new — or everything as known.
response_readiness
enum: not_ready | ready_to_respond | needs_more
Forces: a decision about actionability.
Not every article needs a response. The agent must decide. This is the handoff from thinking to doing.
response_angle
nullable → string
Forces: a position, not a summary.
If the agent is ready to respond, it must pick an angle: agree, dispute, add_nuance, synthesize, reframe. This field prevents neutral summaries disguised as responses.
response_seed
nullable → string
Forces: concrete starting point for a response.
Not a full response. A seed: the first paragraph someone would write. This bridges analysis and action.
citations
array of Citation objects
Forces: every claim traced to a source.
No unsupported assertions. The agent must show its work. This is the trust contract — and it's enforced by a separate schema.
A Real Brief Filled In
From trace run-019e6027. The IBM AI agents article, as the agent saw it.
article_type
industry_shift
core_thesis
The 2025 AI agent narrative is substantially overhyped — current agents are LLMs with basic function-calling bolted on, not truly autonomous systems, and enterprises are not yet ready for the governance, integration, and strategic clarity that real agentic deployment requires.
key_insight
The most consequential barrier to agentic AI adoption in 2025 is not model capability — which IBM experts say is already sufficient — but enterprise readiness: the ability to expose internal APIs, organize proprietary data, and build governance frameworks before agents are deployed at scale.
reading_depth_reason
The key insight is extractable from a single pass — the article's value is in the expert calibration of hype versus reality, but the four narratives follow a predictable structure that yields diminishing returns after the first two sections.
familiarity_level
familiar
response_readiness
ready_to_respond
response_angle
add_nuance
response_seed
The article correctly identifies enterprise readiness as the real bottleneck, but it stops short of naming the specific organizational failure mode: companies are treating agent deployment as a model selection problem when it is actually a data infrastructure and API governance problem. The framing of "agent-ready" enterprises is the right one, but the article never specifies what that readiness actually looks like in practice — which is where the real work is. A response could push on this gap: what does it concretely mean to make your enterprise agent-ready, and why is that harder than it sounds?
The Citation Schema: The Trust Contract
Every claim the agent makes must reference a source. Four source types are allowed. The schema is the falsifiability mechanism: if a claim is wrong, you can trace it to its source and check.
art
article — The claim comes directly from the input article.
Example: "99% of 1,000 enterprise developers surveyed by IBM and Morning Consult are exploring or developing AI agents." Source: article. Traceable, verifiable.
mem
past_brief — The claim comes from a previously stored brief in the agent's memory.
Example: "76% of surveyed executives already view agentic AI as more like a coworker than a tool." Source: past_brief, timestamp included. Cross-referenceable.
web
web_search — The claim was verified through an external web search.
Only appears when the agent chose search_web instead of stop_here. Adds independent verification.
mdl
model_knowledge — The claim comes from the model's training data.
The honest source type. The agent admits it "just knows" this. Flagged differently — these are the claims that carry the highest uncertainty and should be treated as assertions, not facts.
Empty fields speak louder than filled ones.
A field left null is a decision: "I don't know this" or "this doesn't apply." The schema makes ignorance visible. When response_angle is empty, the agent is saying: "I am not ready to take a position." That is information.