Agentic web research needs scoring, not just better prompts

At a glance

What this is: Braintrust evaluates Parallel’s web research agents and finds that deeper tiers and architecture choices change completeness, grounding, and cost in different ways.

Why it matters: IAM and security teams building agentic workflows need the same discipline here, because incomplete sourcing and overconfident outputs create governance blind spots for human and non-human decision systems.

By the numbers:

• Braintrust ran 48 World Cup matchups through six configurations for 288 total runs.
• monolithic-pro and fanout-pro both land around 80% composite quality, but monolithic-pro costs about $0.10 per run against fanout-pro's $0.60.
• The injury specialist task jumps from 39% at base to 95% at core, while the monolithic prompt is still around 51% at core.

👉 Read Braintrust's analysis of agentic web research evaluation and scoring

Context

Agentic web research is a governance problem as much as a tooling problem. Once a system can search, extract, and synthesise sources on its own, the central question becomes whether its outputs are complete, grounded, and auditable enough to trust. That is the relevant lens for agentic AI identity programmes as well, because the research system itself behaves like a non-human decision surface with its own access, evidence trail, and failure modes.

This Braintrust evaluation of Parallel’s web research agents uses structured football matchup research as the test case, but the control lesson is broader. When research is decomposed into machine-generated nodes, relationships, and citations, teams can finally inspect where the system is thin, where it is overconfident, and where a single prompt hides missing evidence. That makes the article most relevant to practitioners designing agentic workflows, review gates, and evidence-based governance.

Key questions

Q: How should teams govern agentic research systems that pull live web sources?

A: Treat them like reviewable automation, not just chat interfaces. Require structured outputs, source provenance, and separate measures for coverage and grounding. A system that can search and synthesize on its own needs controls for evidence quality, traceability, and human approval before its output influences operational decisions.

Q: Why do structured outputs matter for AI governance and reviewability?

A: Structured outputs make it possible to score specific properties such as completeness, citation coverage, and schema utilisation. That matters because prose alone hides failure modes. Once research is stored as nodes and edges with evidence records, reviewers can inspect what was found, what was missed, and whether the answer is actually supported.

Q: What breaks when agentic systems produce complete-looking answers without grounding?

A: Teams get a false sense of reliability. The output may appear rich, yet the supporting evidence may be thin or uneven, which makes later decisions hard to defend. In practice, the failure is not only missing facts. It is the inability to prove which claims were sourced and which were inferred.

Q: When should organisations use specialist agent fan-out instead of a monolithic workflow?

A: Use fan-out when the task has genuinely separable subdomains that benefit from tighter briefs, such as injuries, access issues, or other narrow evidence sets. Use monolithic workflows when cross-cutting relationships matter more than local depth. The deciding factor is whether the synthesis layer can preserve links across the whole task.

Technical breakdown

Why structured research outputs are easier to govern than prose

A structured research output turns open-ended retrieval into a schema that can be scored, inspected, and compared. Instead of reading a narrative and guessing whether the model found enough support, teams get nodes, edges, timestamps, and evidence excerpts that can be validated independently. That matters because agentic systems often fail quietly: they produce fluent answers with uneven sourcing. In governance terms, the structure is the control surface, not the prose. It creates repeatable checkpoints for completeness, provenance, and schema usage, which are the ingredients needed for reviewable automation.

Practical implication: require machine-generated research to emit structured evidence objects, not just summaries.

How processor tiers change coverage and grounding

The article shows that deeper processor tiers do not just improve fluency, they change how much of the task gets covered and how well it is grounded. A shallow run can fill visible fields while still missing source-backed relationships, while a deeper run can surface more evidence but at higher cost. That split is familiar in security automation: more compute does not automatically mean better governance. The useful question is whether the system is improving the right dimension, such as sourcing density or schema coverage, rather than merely looking more complete.

Practical implication: benchmark tiers separately for evidence quality, not just for output length or apparent completeness.

Why multi-agent fan-out can help narrow tasks and hurt synthesis

Fan-out architecture gives specialist agents narrower briefs, which can improve depth in a focused area like injuries or availability. The tradeoff is that the merged result can become more modular, with weaker cross-domain stitching between entities. In agentic workflows, that is the classic decomposition problem: specialisation improves local recall, but synthesis can lose the connective tissue. For identity and security programmes, the lesson is that task decomposition needs a reconciliation layer that can preserve lineage across sub-tasks, otherwise the final answer may be well-populated but weakly connected.

Practical implication: use fan-out only when you also have a clear synthesis and reconciliation stage.

NHI Mgmt Group analysis

Structured agent outputs create governance value only when evidence quality is scored separately from surface completeness. This article shows why a system can look rich at the field level while still missing the supporting basis for its claims. That is the same failure pattern seen in many AI and automation programmes: coverage metrics flatter the output, while provenance gaps remain hidden. Practitioners need to treat evidence coverage as a first-class control, not a by-product of good prompting.

Agentic research introduces a new kind of control gap that looks like observability debt. When a workflow can search, synthesize, and merge results autonomously, the organisation inherits a dependency on the system’s internal traceability. If the review layer cannot distinguish a well-supported edge from a fluent guess, governance becomes performative. The practical conclusion is that agentic systems need auditability built into their data model, not added after the fact.

Named concept: evidence-grounding drift is the point where output richness outpaces sourced validation. The article demonstrates that schema fill and narrative quality can rise faster than citation grounding, which creates a dangerous illusion of reliability. For identity and AI governance teams, that means the control objective is not just output quality, but keeping the evidence trail aligned with the answer. Practitioners should design review gates around grounding drift, not just correctness.

For identity programmes, the bigger lesson is that agentic systems should be governed like non-human actors with constrained evidence scopes. Once a system is allowed to gather, transform, and reuse information independently, it needs lifecycle controls, access boundaries, and reviewability similar to other non-human identities. The article does not present an identity breach, but it does show the same governance shape: a machine process acting with enough autonomy to require explicit oversight. Practitioners should map these workflows into existing NHI and AI governance models before they proliferate.

What this signals

Evidence discipline will become a core requirement for agentic workflows. As more organisations rely on AI systems to gather and synthesise information, the governance question shifts from whether the system can answer to whether it can prove how it answered. That pushes review controls toward provenance, scoring, and traceable transformations rather than vague confidence in model output.

The practical programme impact is straightforward: teams will need separate assurance paths for output quality and source quality. Where the workflow touches identity-sensitive data, secrets, or privileged access, the inability to demonstrate evidence lineage should be treated as a control failure, not a usability issue.

For practitioners

• Define evidence-quality thresholds for agentic research Set minimum standards for evidence coverage, basis coverage, and schema utilization before any machine-generated research can inform decisions. Use separate thresholds so a polished summary cannot pass on prose quality alone.
• Benchmark architectures against the same task Run monolithic and fan-out patterns on identical briefs and compare their grounded output, not just their final answer. Track where specialist decomposition improves local detail but weakens cross-linking between entities.
• Use tiering as a governance decision, not a preference Treat deeper processor tiers as a costed control choice. Approve higher-tier runs only for research domains where the added grounding materially changes the decision, such as injury-heavy or evidence-dense workflows.
• Add reconciliation checks to merged agent outputs When multiple specialist agents contribute to one result, validate that the merge step preserves sources, edges, and relationships across slices. Missing cross-team links are a sign that synthesis needs stronger review.

Key takeaways

• Agentic research can look complete while still being weakly grounded, so evidence quality must be scored separately from output polish.
• Architecture and processor tier change the balance between depth, coverage, and cost, which makes benchmarking a governance task rather than an engineering preference.
• Security and identity teams should treat machine-driven research as a controlled non-human workflow with provenance, review, and lifecycle oversight.

Key terms

• Evidence Coverage: The proportion of outputs that are backed by supporting source material. In structured agentic research, this measures whether relationships and claims are actually grounded rather than merely generated. High evidence coverage improves auditability, but it still needs to be paired with citation quality and source freshness.
• Basis Coverage: A stricter grounding measure that checks whether a claim is supported by usable excerpts, not just a source reference. It helps distinguish outputs that look sourced from outputs that can be defended. In governance terms, basis coverage is closer to proof than to documentation.
• Schema Utilisation: How fully a model uses the fields, relationships, and specialised attributes available in a structured output. It shows whether the system is taking advantage of the full ontology or only filling the easiest parts. For agentic research, schema utilisation is a proxy for depth and task understanding.
• Evidence-Grounding Drift: The gap that appears when an output becomes richer and more complete-looking without a corresponding increase in source validation. It is a useful governance concept for agentic systems because it captures a common failure mode where fluency outruns traceability and the answer becomes harder to trust.

What's in the full article

Braintrust's full analysis covers the operational detail this post intentionally leaves for the source:

• Run-by-run comparisons across six configurations, including how monolithic and fan-out designs changed the research map.
• Scorer definitions and row-level inspection detail for evidence coverage, basis coverage, and schema utilisation.
• Cost comparisons across processor tiers, including where pro tier justified its extra depth and where it did not.
• The full experimental view of how specialist injury research performed differently from broader generalist prompts.

👉 Braintrust's full post covers the configuration comparisons, scorer definitions, and cost tradeoffs in detail.

Deepen your knowledge

NHI Foundation Level course, the industry's only accredited NHI security programme, covers NHI governance, secrets management, and workload identity for practitioners who need stronger control over machine-driven access. It helps identity and security teams build the governance discipline required for non-human and agentic systems.