Task Shape Matters More Than Industry When Picking a Use Case
Industry labels sound strategic.
Task shape is usually a better predictor of whether an agent will work, how you verify it, and which tools it needs.
Summary
- Prefer classifying work by task shape (research, execution, monitoring, and hybrids) before optimizing for industry verticals.
- Insight: Vertical branding drives demos; shape drives architecture, evals, and failure modes.
- Key Concepts: task shape, research, execution, monitoring, hybrid workflows, verification asymmetry, risk overlay.
- When to Use: Selecting a first agent, rewriting a bloated "industry platform" roadmap, or diagnosing why a vertical pilot stalled.
- Limitations/Trade-offs: Industry still matters for compliance, data residency, and irreversible actions. Shape first, risk second - not risk never.
- Related Topics: four use-case categories, agent-fit decision framework, premature use-case signals.
Foundations
A task shape is the recurring pattern of how work moves from goal to done.
Three shapes cover most agent candidates:
- Research - gather, compare, synthesize, cite.
- Execution - change a system of record (code, ticket, CRM, config).
- Monitoring - watch signals, detect exceptions, escalate or act.
Hybrids exist (research-then-execute, monitor-then-execute), but naming the primary shape keeps designs honest.
Industry is a risk overlay on that shape.
"Healthcare research" and "startup research" share retrieval, synthesis, and citation needs.
They differ on PHI handling, audit, and who may act on the output.
If you start with "we need a healthcare agent," teams often buy vertical kits before they know whether the real job is research, documentation execution, or monitoring.
If you start with "we need multi-source research with citations," you can reuse patterns from any domain and then apply healthcare constraints.
Analogy: choosing a vehicle.
Task shape is the terrain (highway, off-road, city).
Industry is the cargo rules (hazmat, refrigerated, passenger).
You pick the chassis for terrain first, then outfit it for cargo regulations.
Mechanics & Interactions
Each shape implies a default loop and a default success check.
| Shape | Loop emphasis | Typical tools | Primary success check | Common failure |
|---|---|---|---|---|
| Research | fan-out search, read, synthesize | search, browsers, docs | coverage + citation quality | confident wrong synthesis |
| Execution | plan, act, verify, repair | APIs, shell, tickets, editors | system state matches goal | partial writes, wrong target |
| Monitoring | sample, detect, rank, notify/act | logs, metrics, queues | precision/recall on alerts | alert fatigue or missed events |
Research agents tolerate longer latency when quality is high.
Execution agents need tight stop conditions and often human gates on write tools.
Monitoring agents need durable schedules, budgets, and quiet hours - they are closer to ops products than chat demos.
Shape also predicts context growth.
Research fills the window with source text and needs summarization discipline.
Execution fills it with diffs, stack traces, and tool errors.
Monitoring fills it with time-series snippets and incident history.
Treating these as one "industry agent" with a kitchen-sink prompt usually fails for context and tool-selection reasons, not model IQ.
How industry should enter the design after shape is fixed:
- Identify regulated data classes and retention rules.
- Mark irreversible actions and required approvers.
- Add audit fields to every tool call and decision.
- Adjust autonomy down until evals and policy allow a step up.
Notice that none of those steps invent a new task shape.
They constrain the same research/execution/monitoring loop.
Cross-shape products need explicit stage boundaries.
Example: competitive intelligence might monitor public sites daily, then run a research synthesis weekly, then execute a CRM update with human approval.
That is three agent (or workflow) stages with different tools and KPIs, not one vague "sales AI."
Advanced Considerations & Applications
Teams that ignore shape fall into two traps.
Trap A - Vertical theater: decks full of industry icons, tools that barely cover one shape, no eval set.
Trap B - Shape denial: a monitoring problem forced into a chat UX because "agents are conversational," producing expensive, non-durable runs.
A practical intake sequence:
- Write the goal and
done_when. - Tag primary shape (+ secondary if needed).
- List tools required for that shape.
- Only then apply industry risk overlay.
- Only then pick autonomy level and framework.
This order reduces framework churn.
LangGraph, CrewAI, Microsoft Agent Framework, and others can implement any shape.
Shape mismatch shows up as missing tools and unmeasurable success, not as "we chose the wrong logo."
Portfolio strategy also follows shape.
Many orgs should ship one strong execution agent (often coding or ticket ops) and one research agent before investing in always-on monitoring agents.
Monitoring is operationally hard: false positives burn trust, and unattended actions raise blast radius.
| Approach | Strength | Weakness | Best Fit |
|---|---|---|---|
| Shape-first design | Reusable patterns, clearer evals | Less flashy in industry decks | Most first agents |
| Industry-first design | Speaks buyer language early | Hides missing tools/verification | Late-stage vertical packaging |
| Hybrid stages (explicit) | Matches real workflows | More orchestration overhead | Multi-step business processes |
| Single blended agent | One UX entry point | Tool confusion, mixed risk | Only with strong routing + policy |
When stakeholders insist on industry first, still force a shape appendix in the design doc.
If they cannot name research vs execution vs monitoring, the project is not ready for an agent budget.
Common Misconceptions
- "Our industry is unique, so generic patterns do not apply." Regulations and data differ; research/execution/monitoring mechanics still transfer.
- "If we hire domain experts, shape does not matter." Experts improve labels and policy; they do not replace tool coverage or stop conditions.
- "Monitoring is just research on a schedule." Monitoring needs durable jobs, alert quality metrics, and often different autonomy rules.
- "Execution always means fully autonomous writes." Most good execution agents draft, propose, or open PRs/tickets first.
- "Task shape is only for product managers." Engineers use it to size context, tools, and eval harnesses.
FAQs
What is task shape in one sentence?
The structural pattern of work - mainly research, execution, monitoring, or a named hybrid - independent of industry branding.
Why not start with industry?
Industry without shape produces demos that cannot define success. Shape first yields tools and metrics; industry then sets constraints.
How do I classify a mixed workflow?
Name the primary shape that creates most value or risk, list secondary shapes as stages, and avoid a single unbounded agent for all stages.
Is coding a shape or a category?
Coding is a use-case category whose dominant shapes are execution and review (with research for unfamiliar code). Category describes environment; shape describes loop style.
Does shape change the model I should use?
Indirectly. Research may favor long-context or tool-using models; execution may favor strong function calling and repair; monitoring may favor cheap classifiers plus occasional deeper analysis. Verify model choice at build time.
How does shape affect human-in-the-loop design?
Research often needs sampling of citations; execution needs approval on writes; monitoring needs escalation policies and on-call ownership.
Can a script implement a shape without an agent?
Yes. Fixed research pipelines and cron monitors are valid. Use an agent when branching on intermediate results is essential.
What is the fastest way to teach stakeholders about shape?
Take three of their tickets and label each research, execution, or monitoring in a 15-minute workshop. Argument quality jumps immediately.
How does this interact with the four categories?
Categories (coding, knowledge, personal, industry) group environments. Inside each, still pick a shape to design the loop and evals.
Why do vertical AI startups still market by industry?
Buyers search by industry. Engineering success still depends on shape-fit. Use industry for packaging after the shape works.
What metrics match each shape?
Research: citation precision and time-to-brief. Execution: success rate and rollback rate. Monitoring: precision/recall and mean time to acknowledge.
Is planning a fourth shape?
Planning is usually a phase inside research or execution (plan-and-execute architectures). Treat it as a control style, not a separate use-case shape, unless planning is the product.
When does industry override shape advice?
When compliance forbids a tool or autonomy level the shape wants. Keep the shape, reduce permissions, and add human gates rather than inventing a mystical new pattern.
Related
- The Four Broad Categories of AI Agent Use Cases - environment buckets to combine with shape
- Use Cases for AI Agents Basics - practice scoring shape on real tasks
- A Decision Framework: Is This Problem a Good Fit for an Agent? - full go/no-go workflow
- Why Industry Context Changes an Agent's Risk Profile - when the risk overlay dominates
- How Knowledge-Work Agents Differ from Coding Agents - category contrast driven by shape
Stack versions: Pins from the category manifest (verify at build): OpenRouter (~315+ models, July 2026 pricing/fees); LangGraph 1.0+; CrewAI 1.14+; Microsoft Agent Framework 1.0; Vercel AI SDK 6; Pydantic AI (latest); LlamaIndex (latest); OpenAI Agents SDK (latest + MCP); MCP (Linux Foundation governance); A2A (HTTP+SSE+JSON-RPC 2.0); Solana
@solana/web3.js+@solana/spl-token.