The irreversible costing shift in software engineering

For most of software history, delivery cost was dominated by the labor of writing and debugging code. In 2025–2026 that assumption broke. Agents made the mechanical loop (draft, run, fix, repeat) dramatically cheaper and massively parallel. Product sense, verification, and system design stayed expensive in human hours.

Call that the costing shift. It is irreversible for the same reason cloud hardware got cheaper. Capability and unit economics both point the same way. Organizations that still budget engineers primarily as authors of lines will mis-allocate the most limited resource they have, which is people. Builders still matter; the job shifts toward judgment, system design, and the fleets those builders command.

Limited human hours sit downstream of the agent loop. When agents absorb draft→test→fix work, typing stops setting the schedule. The queue moves to work that still requires judgment, including review, specification, coordination, and coherence.

Human hours stay the limited input. According to Figure 1, implementation falls from 48% to 12% of human attention (−75%), while review and verification rise from 15% to 34% (+127%). Industry telemetry tracks the same pressure: higher PR volume, larger diffs, and longer review cycles under high AI adoption^[5][9]. When agents absorb draft and patch work, the schedule is set by judgment capacity rather than typing speed. Agents reweight limited human hours toward the judgment work that always existed. That reallocation holds only if agents also got much more capable and much cheaper to run. If execution stayed limited, teams would hire more authors and the keyboard would remain the long pole.

In 2024–2026 capability and cost moved in the same direction. Frontier resolve rates climbed, labs reported order-of-magnitude more agent-authored output, and list and platform token prices fell^[18]. The API cost of a strong score still varies widely by model and harness. Stronger models make bulk agent authorship possible. Lower unit costs make that authorship affordable at fleet scale. Capability, volume, and price rewrite the delivery budget only when they reinforce one another.

According to Figure 2, the frontier envelope climbs from 22% resolve rate in early 2024 to 96.2% by mid-2026 (+337%), then packs several models into the same high band near the ceiling. The line is the highest resolve rate at each date. Lower dots show capability became a pack race across labs.

That pattern supports the costing thesis. Agent execution quality is abundant relative to typing-era limits. When the frontier keeps climbing and alternatives stack on the same calendar, organizations that still budget human attention as if coding were the long pole will underfund review, specification, and harness work, the limited inputs that now set the schedule.

Capability keeps climbing while human review capacity does not scale the same way. Not every release is a new peak, and vendor harnesses differ from public Mini-SWE runs, yet agent execution is more than a temporary spike. Independent time-horizon measures show agent task length roughly doubling on a short cycle, the same direction as the resolve-rate ladder^[21].

According to Figure 3, the four tiles move as one direction of travel. Calendar time falls by about 10× in the OpenAI harness report, Claude-authored merges reach 80%+, merge volume per engineer is about 8× higher than 2024, and a token-cost sample drops about −67% year over year. Treat the multipliers as quantity proxies from labs dogfooding their own tools, not a universal productivity KPI. The joint pattern still matters.

Together with the resolve-rate ladder, that pattern is why reallocation of human attention sticks. When retries and bulk authorship get cheaper, process waste shows up as limited human hours on review, intent, and orchestration rather than calendar typing time. Suite cost against resolve rate then decides whether the leverage stays affordable at fleet scale.

Volume and token cost tell half the economics story. The other half is whether a strong score is cheap enough to run as a default fleet path with model and harness together, not resolve rate alone. That cost–capability surface is the real fleet decision.

According to Figure 4, the scatter is a fleet decision surface. Up is better and left is cheaper. On Verified Mini-SWE, systems can sit near 93–96% resolve rate while unit economics diverge: GPT-5.6 Luna at $0.21/task is about −82% cheaper per task than GPT-5.6 Sol at $1.15/task for 93.0% vs 96.2% resolve. DeepSWE pulls the cloud down and spreads dollars further. Harder work separates a strong score from a package cheap enough to run at scale.

That spread is industry evidence for the costing shift. Capability without unit economics is incomplete. The hard judgment is which model–effort–harness package stays leftward for default automation. Teams that chase headline resolve % alone overpay on volume loops and underfund the human review and design work those loops create, and that reallocation of human hours is the irreversible move.

The frontier climbed and lab volume exploded, but model and harness selection still determines whether that leverage is economically sustainable^[10][43]. Once that economics is real, org design has to change. Leave the old headcount model in place and the bottleneck simply moves into review queues and coordination debt.

When execution is bulk and judgment is limited, headcount plans and rituals built for line authorship miss the queue. An AI-native engineering org flips the default. Agents build. Humans keep the work that compresses poorly, such as product sense, architecture legibility, eval design, taste, and the harnesses that make fleets trustworthy.

According to Figure 5, the limited input flips. When typing is limited, labor owns the schedule, code is the long pole, heavy planning feels rational, and review tracks human authorship. When execution is bulk, judgment is limited, verification sets the calendar, evals and harnesses buy leverage, and review tracks agent throughput unless the org redesigns it. The operating model follows that economics. AI-native teams treat agents as default builders and keep humans on intent, verification, and harness design. “Code wins” over ceremony, overnight agents come with harder review, and smaller cores own more surface^[1][2][4]. When overnight agents open a flood of pull requests, the review queue sets the calendar, not typing time. Teams that fold failures into harness rules keep the leverage; teams that only add more authors re-create the old bottleneck in a new shape.

The pattern is leaving the labs. Anthropic × NEC put Claude Code in front of tens of thousands of employees^[23]. New roles such as loop designers, eval owners, and fleet operators matter more than planning theater^[22]. Those roles earn their keep by designing how work is discovered, dispatched, and verified at scale.

When typing is bulk, the high-leverage move is designing systems that prompt, verify, and repeat without a human in every turn.

A loop is a recursive, goal-directed system that discovers work, dispatches agents, verifies, records state, and repeats. Mid-2026 made that shift concrete from three directions at once^[24]. Boris Cherny, creator of Claude Code, frames the job as writing loops that prompt Claude rather than prompting turn-by-turn^[24][25]. Peter Steinberger, creator of OpenClaw, states the same reframe for high-throughput OSS as designing loops that prompt your agents^[26]. Andrew Ng, founder of DeepLearning.AI and co-founder of Coursera, lifts loop engineering to product cadence with three development loops at minutes, hours, and days^[27][28][29]. The skill moves from prompt craft to system design. Ng’s three product-development loops put those cadences on one map.

According to Figure 6, product work runs on three nested cadences. Agentic coding sits on the order of minutes, developer feedback on the order of hours, and external feedback on the order of days, with product specs and developer vision as shared nodes. Those cadences need runtime primitives before they become a daily loop. Across Codex-style apps and Claude Code the shared stack includes scheduled automations and goals, git worktrees for parallel isolation, Skills (SKILL.md) as progressive project knowledge, MCP/connectors for real tools, sub-agents so maker ≠ checker, and durable state on disk (markdown, tickets, AGENTS.md) because models forget^[24]. A typical daily loop runs as a short pipeline.

According to Figure 7, a daily loop is Triage → Isolate → Draft → Verify → Ship, with Human on exceptions and merge only. That pattern stays safe only when judgment stays in the loop. Loops amplify judgment, good or bad. Stop reading output and you accumulate comprehension debt^[30]. Verification still ends with a human who owns the merge. And a loop is only as good as the environment it runs in, including docs, tools, tests, hooks, and memory that agents can actually use.

A loop without a harness is chat with extra steps. What people experience as “the agent” is the model plus everything wrapped around it. Agent = Model + Harness.

Concretely, a harness includes the durable wrap around the model^[4][22][31]:

According to Figure 8, the coding agent is only half the system. The human steers guides and sensors; feedforward and feedback close the loop into the agent; durable state sits outside the context window so failures can ratchet into permanent harness rules. Agent = Model + Harness is the architecture, not a slogan.

OpenAI’s experiment treats the repository knowledge as system of record. A short AGENTS.md is the table of contents (roughly 100 lines), not the encyclopedia. Structured docs/ holds design docs, execution plans, product specs, references, and quality grades. Progressive disclosure keeps the map in context and the depth on disk. Per-worktree bootable apps and strict linters whose errors include remediation instructions make the agent runnable^[4].

Multi-hour agent runs need the same structure at longer horizons, with clear stop conditions, durable state, and feedback that keeps unattended work on track^[32]. Conference practice adds hooks (red squigglies for agents) and a ratchet mindset. Every observed failure becomes a permanent harness rule^[2][31].

According to Figure 9, short AGENTS.md is the map injected into context, and structured docs/ is the encyclopedia on disk. Progressive disclosure means the agent opens only the branch it needs. That layout also decides what the model re-reads on every turn. System prompts, tool descriptions, skill files, and long agent maps land in the context window before the task does. Dump the full wrap on every step and tokens burn, useful task context gets crowded out, and the prompt prefix changes enough that platforms cannot reuse work. A production harness designs that load on purpose.

Context efficiency is harness engineering under a cost meter. At Code with Claude 2026, Brad Abrams (Claude Platform) put a production floor on that meter. Aim for at least an 80% prompt-cache hit rate before other agent optimizations. Mature product harnesses already clear a higher bar. Cursor, Replit, Perplexity, and Claude Code itself run cache rates in the 90s^[2]. Cached tokens are cheaper and faster, and they do not count against rate limits. High hits mean the wrap is stable and lean enough that strong models and frequent agent retries can stay the default production path.

Day to day, the costing shift shows up as operating habits. Invest people and budget in the steps that still set the schedule. The durable investments are loops that keep agents moving without a human on every turn, harnesses that make those loops safe and reusable, and verification capacity that scales with agent throughput. More authors alone re-create the old bottleneck. The habits below turn that investment stance into workflow, harness, and org practice.

Adopt the habits with the same caution used for the evidence itself. Conference practice, harness write-ups, and loop engineering notes converge on the same investment stance; they do not guarantee every tip fits every codebase.

Four questions still set the research and operating agenda.