Bounded vs Unbounded Tasks
frameworkagentsautonomyhuman-in-the-loop
Bounded vs Unbounded Tasks
A practical framework from Praveen Akkiraju for deciding where agent autonomy is realistic and where humans must stay in the loop.
The dichotomy
| Bounded tasks | Unbounded tasks | |
|---|---|---|
| Output specifiable? | Yes — clear right answer | No — many valid answers, much context |
| Verifiable? | Yes (math, tests, types) | Hard — judgment-based |
| Examples | Coding, language migrations, security vuln remediation, doc generation, test generation | Supply chain across SKUs/geos/suppliers; complex sales; novel product strategy |
| Path to autonomy | Fast — RL works, agents converge | Slow — depends on harness quality and domain context |
| Today's reality | Can be ~80–100% autonomous (per Blitzy data) | Human-in-loop necessary; the dial moves slowly |
Why this framework matters
It collapses several debates into a single axis:
- Where to start with agents → bounded first, always
- Build vs buy → bounded workflows are more standardized, often buy; unbounded are domain-specific, often build (see Build vs Buy (Agents))
- Pricing models → bounded tasks are easier to price as fractional FTEs; unbounded resists clean unit pricing
- Hiring → for unbounded work, harness-design skill matters more than coding speed
Three forces shaping the Human in the Loop dial (per Praveen)
- Workflow nature — bounded ↔ unbounded
- Regulatory/compliance environment — healthcare, finance, legal need human signoffs even for bounded tasks
- Harness quality — see Harness (LLM Agents)
The IC-level cousin: ARR Framework
You're Not Behind (Yet) Learn AI Agents (theMITmonk) gives the personal-productivity version of the same axis. ARR (Autonomous / Recurring / Reviewable) is a one-line gate for "should I make this an agent or use a prompt?" — applicable to a single knowledge worker's workflow, not an enterprise deployment.
The two frames compose:
- ARR → answers "is this an agent shape at all?"
- Bounded/Unbounded → if yes, "how aggressively can it be autonomous?"
Sandeep's Narrow Agents thesis (the agents that win are obsessively narrow) is also bounded-flavored: narrow ≈ bounded + commercially specific. Adds a founder/operator lens to what Praveen frames at the enterprise-architecture level.
Triangulation
- Autonomous Software Development with Blitzy (CXOTalk) — GNP picked highly bounded use cases first (Java upgrade, frontend modernization, vuln remediation). Got 80–100% autonomy. Expected outcome under this framework.
- Andrej Karpathy on Agentic Engineering (Sequoia AI Ascent) — Karpathy's "verifiability" framing maps to bounded. "Coding is at AGI" per labs because coding is the most verifiable common task.
- Boris Cherny on Coding Is Solved (Sequoia AI Ascent) — Boris's 100% autonomy in TypeScript+React is the maximum-bounded extreme.
Sources
- Agentic AI in the Enterprise (Praveen Akkiraju, CXOTalk) (canonical)
- Autonomous Software Development with Blitzy (CXOTalk)
- You're Not Behind (Yet) Learn AI Agents (theMITmonk) (the IC-level ARR cousin; the narrow-ownership thesis)