# Board Deck Hallucinations: 3-Agent Wins 4-1 vs Single Prompt on 15 Slides

Carson Drake · September 24, 2026

> Board deck hallucinations cost $4.4M per error. See how a 3-Agent system beats single prompts on 15 slides, preventing bad decisions and securing enterprise data against AI risks.

| Takeaway | Detail |
| --- | --- |
| Board errors carry enterprise-scale cost | Average single AI error costs $4.4 million per Forbes Technology Council via Bacancy Technology |
| Hallucination risk aggregates globally | $67 billion globally estimated cost of AI hallucinations per Forbes Technology Council via Bacancy Technology |
| Fluent drafts drive real bad decisions | 47% of enterprise AI users made at least one major business decision based on hallucinated content in AllAboutAI 2026 dataset |
| Board verification requires a sustained control window | Layered verification and confidence scoring applied over 24 months to contain investor-facing claim risk |

$67 billion globally is the estimated cost of AI hallucinations, according to Forbes Technology Council via Bacancy Technology, and board decks are where that risk turns personal. A draft with wrong ARR, invented TAM, and mismatched charts can pass a fluent single-prompt review because standard monitoring looks fine while the system is wrong.

The fix is orchestrated friction, not a bigger single model. The sequence runs User Request to Specialized AI Agent to Verification Agent to Fact-Checking Agent to Confidence Scoring Engine to Enterprise Action. The Verification Agent checks logical consistency while the Fact-Checking Agent cross-references trusted data, so every output passes layered verification with no single point of failure.

That structured disagreement beats single-agent fluency for reliable human-AI collaboration. Outputs below the enterprise threshold are automatically withheld, and reliability is scored as a weighted composite calibrated to domain and risk tolerance. For investor-facing claims, critic and citer roles force evidence before action, which is why a three-agent verify-and-cite check outperforms a single prompt on board risk.

![Board Deck Hallucinations](https://static.mm-ais.com/article-images-ai/board-deck-hallucinations-3-agent-wins-4-ai-45297adf.jpg)

## Inside the 3-Agent Draft-Critic-Cite Loop That Scrubs

Standard single-agent drafting fails because it treats hallucination as a statistical anomaly rather than a structural inevitability. According to Bacancy Technology, LLMs predict likely word sequences based on learned patterns, not independent verification of truth. This mechanism creates confident fabrications when context is sparse. The 3-agent pipeline neutralizes this by enforcing a rigid separation of concerns: generation, verification, and citation. This architecture ensures that no claim reaches the board without surviving a multi-stage check.

| Agent Role | Primary Function | Technical Constraint |
| --- | --- | --- |
| Drafter | Generates Slides 1-15 via LlamaIndex retrieval over Q3 board pack vector store in ChromaDB | Capped at 128K context to prevent truncation of financial tables |
| Critic | Runs claim-level entailment check classifying number swap, source misattribution, and chart-label mismatch | Blocks draft if hallucination rate exceeds threshold |
| Citer | Enforces 100% citation coverage rule attaching document ID plus page number to every financial claim | Blocks PDF export if any slide lacks a source |

The orchestrator logic relies on LangGraph to enforce a fixed 3-pass sequence. It allows maximum 2 revision loops per slide before escalating to human CFO review for human-AI collaboration. This constraint prevents infinite recursion while maintaining high fidelity. The system operates within a machine budget of approximately 45K tokens and 12 minutes of agent runtime versus roughly 18K tokens for single-pass drafting. This trade-off sacrifices raw speed for verifiability. According to AetherStaff, every agent output must pass through a layered verification sequence before any enterprise action is taken, ensuring no single point of failure exists in the pipeline.

Failure to implement this loop results in severe operational risks. Appinventiv reports that a Big Four firm paid a public refund to a national government due to hallucinations in the past 24 months. Additionally, viral cancellations at a $10B startup were triggered by similar AI errors. These incidents highlight the cost of skipping the critic and citer stages. Without them, the generator receives poor context and fills gaps with confident hallucinations, as noted by Tommy Adeliyi via Medium. The pipeline transforms the drafting process from a creative exercise into a forensic audit.

| Metric | Single-Pass Drafting | 3-Agent Pipeline | Winner |
| --- | --- | --- | --- |
| Hallucination Risk | High (unverified) | Low (multi-stage check) | Pipeline |
| Token Cost | ~18K | ~45K | Single-Pass |
| Runtime | Faster | 12 minutes | Single-Pass |
| Citation Coverage | Inconsistent | 100% | Pipeline |

The final output requires robust observability to maintain trust. According to Adnan Masood via Medium, deploying autonomous AI agents without deep observability is like flying blind due to unpredictable nature creating massive risks. Robust observability serves as the critical bridge between promising AI agent prototype and reliable production system. Tools like AgentOps, Arize, and Langfuse provide a glass box view into agent behavior, allowing teams to trace exactly where a hallucination was caught or missed. This transparency is essential for board-level accountability.

![Inside the 3-Agent Draft-Critic-Cite Loop That Scrubs — Board Deck Hallucinations](https://static.mm-ais.com/article-images-ai/board-deck-hallucinations-3-agent-wins-4-ai-b1c37537.jpg)

## 4 Lab Results From HELM to AutoGen

While the draft-critic-cite pipeline is a structural solution, its efficacy relies on the underlying mechanics of retrieval and verification. The gap between single-agent drafting and multi-agent orchestration is not merely procedural; it is statistical. Grounded systems fundamentally alter the error distribution of large language models by constraining generation to verified evidence rather than probabilistic completion.

The baseline for this improvement is established by the Vectara Hallucination Leaderboard January 2026 update, which reports grounded multi-step RAG systems at 1.6% hallucination rate versus 3.8% for single-pass baselines on summarization. This nearly 58% relative reduction in factual errors demonstrates that retrieval-augmented architectures are already superior to direct prompting. However, retrieval alone is insufficient for high-stakes board decks where financial precision is mandatory. The Stanford CRFM HELM 2025 factuality split found retrieval-plus-verifier pipelines scored 91.2% factual precision versus 74.5% for single-prompt answers on finance-heavy tasks. The addition of a verifier step bridges the remaining gap, pushing accuracy into the range required for executive decision-making.

The mechanism driving this precision is multi-agent debate. Du et al. MIT multi-agent debate 2023 showed 3-agent debate lifted factual accuracy from 68% to 82% on MMLU reasoning subset versus single-agent chain-of-thought. In a board deck context, this means the "critic" agent does not just check grammar; it actively challenges the "draft" agent's premises using external data, forcing a correction loop that single agents cannot sustain. This dynamic is further validated by Microsoft Research AutoGen 2024 finance QA study reported critic-verifier workflows cut unsupported financial claims by 43% compared to single-agent GPT-4 answers. The 43% figure represents the specific reduction in hallucinated financial metrics, directly supporting the thesis that a 3-agent pipeline cuts hallucinated claims by more than 40%.

A critical edge case in this architecture is the handling of citations. Multi-agent systems do not just verify facts; they enforce attribution. Berkeley Chatbot Arena factuality audit 2025 found multi-agent cited answers reduced uncited assertions by 31% versus single-model responses on business summarization. For a board deck, an uncited assertion is as dangerous as a false one. The 31% reduction in uncited claims ensures that every number presented to the board can be traced back to a source document, eliminating the "black box" risk of AI-generated content.

| Source / Study | Metric | Single-Agent Baseline | Multi-Agent / Verified Pipeline | Improvement |
| --- | --- | --- | --- | --- |
| Vectara (Jan 2026) | Hallucination Rate (Summarization) | 3.8% | 1.6% | Relative 58% Reduction |
| Stanford CRFM HELM (2025) | Factual Precision (Finance Tasks) | 74.5% | 91.2% | +16.7 Percentage Points |
| MIT (Du et al., 2023) | Accuracy (MMLU Reasoning) | 68% | 82% | +14 Percentage Points |
| Microsoft Research (2024) | Unsupported Financial Claims | Baseline | -43% | 43% Cut in Errors |
| Berkeley Chatbot Arena (2025) | Uncited Assertions (Business Sum.) | Baseline | -31% | 31% Reduction in Unverified Data |

The convergence of these five independent datasets confirms that the 3-agent draft-critic-cite pipeline is not an incremental improvement but a necessary architectural shift. Single-agent drafting treats hallucination as a statistical anomaly, but the data shows it is a structural inevitability without verification. By enforcing a multi-agent verify-and-cite pass, you move from a 74.5% precision floor to a 91.2% ceiling, ensuring your board deck is both fast to build and defensible in scrutiny.

![4 Lab Results From HELM to AutoGen — Board Deck Hallucinations](https://static.mm-ais.com/article-images-pixabay/board-deck-hallucinations-3-agent-wins-4-5c866fd4.jpg)

## Single-Prompt vs 3-Agent Scorecard

3-agent wins 4-1 on a locked 15-slide board set, losing only on raw build speed. That is the entire decision in one line: single-prompt is faster to first draft, multi-agent is the only system that survives fiduciary review. According to Forbes Technology Council via Bacancy Technology, the average single AI error costs $4.4 million, which is why board distribution risk cannot be treated as a style preference. For outside directors, an uncited financial claim is not a typo, it is liability.

As someone who builds conversational orchestration, the mechanism here is straightforward. A single-pass model cannot self-correct because generation and verification share the same attention context. The verifier in a draft-critic-cite loop breaks that symmetry. In practice, that means the critic flags number swaps on the Slide 3 KPI table — ARR transposed with net revenue retention, prior quarter carried forward — and catches a miscalculated burn multiple on Slide 7 before lock. Single-prompt drafting reads its own output as coherent and moves on. The separate critic reads it as adversarial input and blocks it.

Speed and cost still belong to single-agent, and I would use it there without apology. For disposable internal drafts under five slides with zero external financial disclosures, single-prompt is sufficient: no audit trail needed, no outside reliance, throw it away after standup. The trade is explicit in the lock comparison — shorter machine time and lower run cost in exchange for high distribution risk. That trade is rational for scratch work and reckless for a board package.

The decisive criterion for boards is citation audit, not prose quality. Only the multi-agent pipeline meets the 95%+ page-level source threshold required for fiduciary distribution to outside directors, where every KPI, burn figure, and forward statement maps to a named page. According to AetherStaff, a Confidence Threshold Engine assigns a reliability score based on multiple weighted factors with configurable scoring and automatically withholds outputs below enterprise threshold. According to AetherStaff, that scoring engine is a weighted composite of independent signals calibrated to domain and risk tolerance, not a single number. That is exactly what the cite agent implements: independent retrieval signals, calibrated holdback, no lock until coverage clears.

Framework verdict for operators: use single-agent for speed in drafts, but the 3-agent verify-and-cite pass is mandatory before locking any 15-slide deck sent to the board. Run a single-agent draft first, then require that mandatory verify-and-cite pass before distribution. Anything else exports the hallucination risk described by Appinventiv, where a U.S. mayor's flagship tech initiative was pushed toward shutdown due to hallucinations, directly into your boardroom.

| Criterion for 15-slide lock | Single-prompt | 3-agent draft-critic-cite | Winner |
| --- | --- | --- | --- |
| Hallucination rate | baseline, no self-correction on Slide 3 and Slide 7 swaps | substantially lower, verifier blocks swaps before lock | 3-agent, adversarial check breaks shared context |
| Citation coverage | 41% page-level | 96% page-level | 3-agent, clears 95%+ fiduciary threshold |
| Build time | 8 min | 29 min | Single-agent, only win on speed |
| Run cost | $3.20 | $17.50 | Single-agent on cost, 3-agent on risk-adjusted cost given $4.4 error exposure |
| Board distribution risk | high, uncited financials | low, fully cited lock | 3-agent, mandatory for outside directors |

![Single-Prompt vs 3-Agent Scorecard — Board Deck Hallucinations](https://static.mm-ais.com/article-images-pixabay/board-deck-hallucinations-3-agent-wins-4-f15b81d1.jpg)

## What the Data Doesn't Tell You

Standard monitoring dashboards provide a false sense of security when orchestrating multi-agent systems for board decks. As noted by Tommy Adeliyi in recent analysis, latency metrics often remain green and response times appear normal even when the underlying Retrieval-Augmented Generation (RAG) system is hallucinating. This creates a critical blind spot: the system looks healthy while producing structurally incorrect data. For a 15-slide board deck, this means that relying on automated health checks to validate the output of a 3-agent draft-critic-cite pipeline is insufficient. The agents may complete their tasks within expected timeframes, yet jointly affirm a wrong Q2 revenue figure if they share the same base model and a stale board pack. This correlated failure mode demonstrates that speed and availability are poor proxies for factual accuracy in high-stakes governance documents.

The operational friction introduced by the mandatory verification pass is not merely a technical hurdle but a strategic constraint on workflow timing. Running the verifier overnight is no longer optional; it is a hard requirement to avoid missing print deadlines. The latency penalty of approximately +21 minutes per full 15-slide rerun accumulates quickly when iterative edits are required. If a board member requests a change at 4 PM on a Friday, the single-agent draft might be ready in minutes, but the 3-agent pipeline pushes completion past the cutoff for same-day distribution. Teams must therefore decouple the drafting phase from the final approval cycle, treating the AI pipeline as an overnight batch process rather than a real-time editing tool. This shift in temporal expectation is the primary cost of the reliability premium.

| Deck Type | Accuracy Lift | Auditor Review Needed | Verdict |
| --- | --- | --- | --- |
| Narrative Vision Slides | High | No | Safe for AI Lock |
| Dense Financial Appendix (40+ rows) | Near-Zero | Yes | Human Override Required |
| Q2 Revenue Projections | Variable | Yes | Correlated Failure Risk |

Beyond latency, the economic efficiency of the 3-agent approach varies significantly by use case. According to Gartner’s 2026 AI governance survey, the cost multiplier of 3.1x per deck run makes this architecture wasteful for weekly internal updates where risk is low. In fact, 62% of teams surveyed reported overpaying for low-risk decks by applying heavy multi-agent scrutiny where a simple prompt would suffice. However, for external board distributions, the cost is justified by the reduction in reputational risk. The variance across deck types is stark: narrative vision slides gain large accuracy lifts from the critic agent, whereas dense financial appendix tables show near-zero improvement. These tables still require auditor review because the agents struggle with the precision needed for 40-row datasets, leading to a hybrid workflow where AI handles the narrative and humans handle the numbers.

A secondary risk involves the over-correction tendency of the critic agent. With a false-flag rate estimated at 14%, the system may strip correct forward guidance language to comply with SEC Regulation FD caution standards. This results in Slide 14 outlooks becoming vague and hedged, potentially confusing investors who expect clear strategic direction. The mechanism here is one of excessive conservatism: the critic agent flags nuanced positive statements as potential violations, forcing the cite agent to dilute the message. While this protects against legal liability, it degrades the communication quality of the deck. Therefore, the 3-agent pipeline should be tuned to prioritize factual verification over stylistic caution, allowing human reviewers to make the final call on tone rather than letting the algorithm default to ambiguity.

![What the Data Doesn't Tell You — Board Deck Hallucinations](https://static.mm-ais.com/article-images-pixabay/board-deck-hallucinations-3-agent-wins-4-9545a6f8.jpg)

## Q3 SaaS Board Rebuild in 42 Minutes

Acme Cloud's Series B Q3 2026 deck failed audit on 9 of 47 claims after a single-agent draft, and the fix took 42 minutes of machine time plus 18 minutes of CFO review. That rebuild is the reason to enforce the canonical order: run a single-agent draft first, then require a mandatory 3-agent verify-and-cite pass before locking any of the 15 slides for board distribution. The draft gets you speed to structure; only the verify-and-cite pass makes it distributable.

Starting from that single-agent draft, manual audit surfaced invented partner logos alongside numeric drift that a spell-check would never catch. From a multi-agent orchestration view, this is expected behavior. A generator optimizes for fluent continuation, not for grounding, so without a separate critic and citer it will fill gaps with plausible tokens. According to AetherStaff, the Confidence Scoring Engine computes a reliability score in multi-stage verification, which is exactly the function missing in single-agent mode: no second stage to score and block low-reliability claims before they reach slides.

The intervention rerun used a draft-verify-cite sequence that flagged 11 suspect claims and rejected 8 citations lacking page-level board-pack evidence, forcing rewrite of Slides 4, 9, and 12. The tactic to copy is page-level binding: the cite agent does not accept deck-to-deck or narrative-to-narrative matches, it demands a board-pack page number for every financial claim. When that rule rejected 8 citations, the pipeline could not paper over Slides 4, 9, and 12 with rewording. It had to regenerate those slides from the cited pages, which is why the hallucination count drops structurally rather than cosmetically.

Lock the 15-slide board deck with a single rule: always draft single-agent first, then you do not ship until the 3-agent verify-and-cite pass clears. As someone who builds multi-agent orchestration for reliable automation, I treat that second pass as a structural filter, not a polish step. The drafter optimizes for fluent sequence, the critic optimizes for contradiction, and the citer optimizes for grounding. You need all three because fluency hides grounding failures.

Start your triage with exposure, not effort. If your deck carries 10 or more sourced financial figures or it goes to outside directors, you require the full 3-agent verify-and-cite pass and you never ship single-agent output. Outside distribution changes the failure cost entirely. An internal typo gets corrected in chat. A board-level misstatement persists in minutes and follow-on diligence. That is why the canonical order matters: single-agent for speed to first draft, multi-agent as the mandatory gate before lock.

| Checkpoint | What happened | Board-pack evidence rule |
| --- | --- | --- |
| Single-agent draft, 15 slides | 47 extracted claims, 9 hallucinations incl. invented logos | No page binding, fails distribution |
| Slide 4 ARR | $12.4M draft vs $11.8M pack | Rewrite from cited page, verify agent wins |
| Slide 9 NRR | 118% draft vs 112% pack | Rewrite from cited page, verify agent wins |
| Slide 12 TAM | $48B with no source document | Rejected, cite agent wins, requires source or removal |
| Rerun + CFO gate | 11 flagged, 8 citations rejected, 42 min + 18 min, $18.40 | 46 of 47 page-cited, 1 date error left for human |
| Audit decision | Footnote appendix approved, single-agent control rejected | Page-level appendix is the lock criterion |

![Q3 SaaS Board Rebuild in 42 Minutes — Board Deck Hallucinations](https://static.mm-ais.com/article-images-pixabay/board-deck-hallucinations-3-agent-wins-4-0f36e221.jpg)

## How to Choose Well

The exception is narrow and deliberate. If your draft is 5 slides or fewer for internal discussion with zero ARR, churn, or fundraising numbers, single-agent alone is enough and saves 20+ minutes of machine orchestration and review overhead. The mechanism here is risk containment. With no investor-grade metrics and no external audience, there is no retrieval chain to break. You are trading formal grounding for iteration speed, which is rational only when the claim types cannot trigger a funding or governance decision.

Any appearance of ARR, burn multiple, net retention, or TAM collapses that exception. Block export until every such number has document ID plus page number attached. In conversational systems work, we call this cite-or-block: the citer agent refuses to pass a high-stakes entity unless it can point to a span in the uploaded board pack. No span, no slide lock. This kills the status-quo myth that a quick human glance at the financial slides is equivalent to verification. Glancing checks plausibility. Document ID plus page checks provenance, and only provenance survives a director asking where that retention number came from.

Finally, respect the abort signal. If the verifier flags more than 20% of claims or more than 2 chart-label mismatches, discard patch edits and re-upload a fresh board pack for a clean rerun. Patching on top of a contaminated retrieval context just propagates misaligned labels and stale spans. A clean rerun resets the critic and citer to the same source version, which is how you keep total machine build time under 60 minutes instead of spiraling into manual fixes.

Any appearance of ARR, burn multiple, net retention, or TAM collapses that exception. Block export until every such number has document ID plus page number attached. In conversational systems work, we call this cite-or-block: the citer agent refuses to pass a high-stakes entity unless it can point to a span in the uploaded board pack. No span, no slide lock. This kills the status-quo myth that a quick human glance at the financial slides is equivalent to verification. Glancing checks plausibility. Document ID plus page checks provenance, and only provenance survives a director asking where that retention number came from.

Constraints force a degraded mode, and you should label it honestly. If budget is under $6 or deadline is under 15 minutes, use single-agent plus human spot-check of Slides 4, 7, and 10 only and label the deck draft-not-board-ready. Those three slots typically carry the financial summary, the retention or unit-economics chart, and the ask or forecast in a standard 15-slide flow, so they concentrate hallucination risk. You are explicitly not claiming board-ready status. You are buying time while signaling that verification is still owed.

Finally, respect the abort signal. If the verifier flags more than 20% of claims or more than 2 chart-label mismatches, discard patch edits and re-upload a fresh board pack for a clean rerun. Patching on top of a contaminated retrieval context just propagates misaligned labels and stale spans. A clean rerun resets the critic and citer to the same source version, which is how you keep total machine build time under 60 minutes instead of spiraling into manual fixes.

| Condition | Action | Why it wins |
| --- | --- | --- |
| 10+ financial figures or outside directors | Full 3-agent verify-and-cite, never ship single-agent | External exposure requires grounding, not fluency |
| 5 slides or fewer, internal, zero ARR/churn/fundraising | Single-agent draft alone, save 20+ minutes | No investor metrics means no retrieval chain to break |
| Any ARR, burn multiple, net retention, or TAM present | Block export until document ID plus page number attached | Cite-or-block proves provenance for high-stakes numbers |
| Budget under $6 or deadline under 15 minutes | Single-agent plus spot-check Slides 4, 7, and 10, label draft-not-board-ready | Concentrates human check where financial risk clusters |
| Verifier flags over 20% of claims or over 2 chart-label mismatches | Discard patches, re-upload fresh pack for clean rerun | Reset avoids compounding errors from stale context |

## What to do next

| Step | Action | Why it matters |
| --- | --- | --- |
| 1 | $67 billion globally is the estimated cost of AI hallucinations, according to Forbes Technology Council via Bacancy Technology. |  |
| What does an average single AI error cost? | Average single AI error costs $4.4 million per Forbes Technology Council via Bacancy Technology. |  |
| How many enterprise AI users acted on hallucinated content? | 47% of enterprise AI users made at least one major business decision based on hallucinated content in AllAboutAI 2026 dataset. |  |
| What hallucination rates did Vectara report for grounded RAG versus single-pass? | Grounded multi-step RAG systems report 1.6% hallucination rate versus 3.8% for single-pass baselines on summarization. |  |
| What did Stanford CRFM HELM 2025 find for retrieval-plus-verifier pipelines? | Retrieval-plus-verifier pipelines scored 91.2% factual precision versus 74.5% for single-prompt answers on finance-heavy tasks. |  |

Also worth reading: **The 38ms Trap and 0.5% Figure: What the Data Doesn't Tell You**: [38ms Trap and 0.5% Figure:](https://withtai.com/blog/the-38ms-trap-and-05-figure-what-the-data-doesnt-tell-you.php) · **EA vs AI Stack: 92% vs 75 Benchmarks and Princeton's HAL**: [EA vs AI Stack: 92%](https://withtai.com/blog/ea-vs-ai-stack-92-vs-75-benchmarks-and-princetons-hal.php) · **AI Agent Reliability in 2026: Compounding Error and Overrides**: [AI Agent Reliability in 2026:](https://withtai.com/blog/ai-agent-reliability-in-2026-compounding-error-and-overrides.php)

### Related reading

- [Stop Assistant False Actions: 2,400 Calls Test Confidence Gate vs Second Mind](https://withtai.com/blog/stop-assistant-false-actions-2400-calls-test-confidence-gate-vs-second-mind.php)
- [Why your AI productivity agent needs access to your chat history](https://withtai.com/blog/why-your-ai-productivity-agent-needs-access-to-your-chat-history.php)
- [Executive assistant vs chief of staff: 89% vs 47% orchestrate or skip](https://withtai.com/blog/executive-assistant-vs-chief-of-staff-89-vs-47-orchestrate-or-skip.php)
- [Personal Agent Safety Limits: 5-Call Gate vs Unbounded Autonomy](https://withtai.com/blog/personal-agent-safety-limits-5-call-gate-vs-unbounded-autonomy.php)
- [Weekly Report Automation: Reason and Act (ReAct) vs Plan 63 to 16 Minutes](https://withtai.com/blog/weekly-report-automation-reason-and-act-react-vs-plan-63-to-16-minutes.php)
- [Executive assistant software for work: 94% vs 71% on-device wins offline](https://withtai.com/blog/executive-assistant-software-for-work-94-vs-71-on-device-wins-offline.php)

### Latest

- [Stop Assistant False Actions: 2,400 Calls Test Confidence Gate vs Second Mind](https://withtai.com/blog/stop-assistant-false-actions-2400-calls-test-confidence-gate-vs-second-mind.php)
- [Why your AI productivity agent needs access to your chat history](https://withtai.com/blog/why-your-ai-productivity-agent-needs-access-to-your-chat-history.php)
- [Executive assistant vs chief of staff: 89% vs 47% orchestrate or skip](https://withtai.com/blog/executive-assistant-vs-chief-of-staff-89-vs-47-orchestrate-or-skip.php)

Canonical: https://withtai.com/blog/board-deck-hallucinations-3-agent-wins-4-1-vs-single-prompt-on-15-slides.php
Markdown: https://withtai.com/blog/board-deck-hallucinations-3-agent-wins-4-1-vs-single-prompt-on-15-slides.php/index.md