Skip to main content
A/KAI Kills What?

Dossier #1 // career autopsy

Moderate exposure

Software Developers

Category: Computer & mathematical · SOC 15-1252.00 · Industry: Software & technology

AI Exposure

35/100
0.0

Low exposure

Last updated

12 Oct 2026

16 hours ago · sources: BLS OES (approx) · O*NET

Evidence

9 sources

11 core tasks shown · 6 secondary hidden

Verdict

Software Developers Moderate exposure: only 4 of 11 core tasks show high task automation potential : exposure 35/100 (low exposure).

1 model with published evidence against this job. Strongest: Claude Opus 4.1 (93/100 capability).

Share this verdict:

Score breakdown

Why this score?

Every component is stored with the score row and can be recomputed from the evidence table.

AI Exposure (composite)35/100

Methodology 2026-09-a · score v1.0 · input hash 2c48de3050aed617

AI capability exposure0/100

Strength of current models on the capabilities this job's tasks require.

Task automation potential47/100

Task-level exposure, weighted by task importance. A component of the headline, not the headline itself.

Observed AI adoption56/100

Evidence that AI is actually being used for this work.

Labour-market signal0/100

Observed changes in postings, employment or earnings.

Evidence strength98/100

Quantity and source quality of the evidence behind the score.

9 evidence item(s) contribute. Calculated 16 hours ago. 12 Oct 2026 UTC.

Overview

Research, design, and develop computer and network software or specialized utility programs. Analyze user needs and develop software solutions, applying principles and techniques of computer science, engineering, and mathematical analysis. Update software or enhance existing software capabilities. May work with computer hardware engineers to integrate hardware and software systems, and develop specifications and performance requirements. May maintain databases within an application area, working individually or coordinating database development as part of a team.

Employment

≈1,610,000

approx. BLS OES

Median wage

≈$133,000

approx. BLS OES

Tasks tracked

17

linked to this occupation

Employment and wage values are approximate reference figures from the BLS Occupational Employment and Wage Statistics programme, stored with an approximation flag. Verify against the O*NET record and BLS before quoting them.

Tasks

Task-level automation exposure

All tasks →

Core tasks only. Exposure describes how much of the task output current AI can produce or substantially assist - not whether the task disappears. Secondary shared-baseline edges are hidden until per-occupation curation is complete.

Killers

Which models affect this job most

Ranked by strongest demonstrated capability. Task and evidence counts come from the attributed evidence trail below.

  1. 01
    Claude Opus 4.193/100

    Anthropic · 1 evidence item

Evidence

Evidence trail

9 attributed item(s). Confidence reflects study design: randomised and production deployments score higher than platform observations, which score higher than projections.

  1. task automationincreasesconfidence 90/100 · magnitude 745 Dec 2024

    In pre-registered experiments with experienced open-source developers, access to AI tools increased the number of tasks completed by about 26% relative to control.

    Complements the Copilot study with a more experienced population and a pre-registered design. Reinforces a high task-automation component for implementation work; no employment effect is measured.

    Source: Cui, Demirer, Jaffe, Musolff, Peng & Zheng — The Effects of Generative AI on High Skilled Work (arXiv:2412.04638) · Updated 5 Dec 2024

  2. task automationincreasesconfidence 88/100 · magnitude 7813 Feb 2023

    In a controlled experiment with 95 developers, those with access to GitHub Copilot completed a specified programming task 55.8% faster than the control group.

    A single, well-specified task in one language. It is strong evidence that routine implementation work speeds up, and weak evidence about whole-job effects, particularly for system design or unfamiliar legacy code.

    tool: GitHub Copilot

    Source: Peng, Kalliamvakou, Cihon & Demirer — The Impact of AI on Developer Productivity (arXiv:2302.06590) · Updated 13 Feb 2023

  3. model releaseincreasesconfidence 85/100 · magnitude 585 Aug 2025

    Claude Opus 4.1 extended sustained autonomous performance on multi-step coding and research tasks, with providers reporting multi-hour task horizons.

    Agentic reliability is the capability that moves automation from single-prompt assistance toward completing whole workflows. It affects multi-step task exposure, not physical work.

    model: Claude Opus 4.1

    Source: Anthropic · Updated 5 Aug 2025

  4. researchincreasesconfidence 82/100 · magnitude 7424 Aug 2023

    Using O*NET task descriptions scored by GPT-4, the authors estimate that about 80% of the US workforce could have at least 10% of their tasks affected by large language models, and about 19% could see at least 50% of their tasks affected.

    This is a *capability and exposure* estimate produced by having a model rate task descriptions — not a measurement of employment change. It is the most-cited exposure figure in the field and is included here as an exposure input, weighted accordingly.

    Source: Eloundou, Manning, Mishkin & Rock — GPTs are GPTs (arXiv:2303.10130) · Updated 24 Aug 2023

  5. researchdecreasesconfidence 78/100 · magnitude 4513 May 2024

    Estimating that only a small share of tasks are currently cost-effective to automate, the paper projects total-factor-productivity gains of no more than about 0.66% over ten years, with limited aggregate labour-demand effects.

    Included as a counterweight. Exposure — even high task exposure — does not imply large aggregate job loss, because adoption depends on cost, reliability, regulation and complementary investment. This evidence lowers the labour-market component rather than capability.

    Source: Acemoglu — The Simple Macroeconomics of AI (NBER Working Paper 32487) · Updated 13 May 2024

  6. adoptionincreasesconfidence 72/100 · magnitude 541 Apr 2025

    The AI Index reports continued increases in the share of organisations using generative AI in at least one business function, alongside rapid benchmark gains on reasoning and coding tasks.

    Survey-based adoption evidence across firms. Useful for the adoption component and for interpreting why capability and adoption diverge by occupation.

    Source: Stanford HAI — Artificial Intelligence Index Report 2025 · Updated 1 Apr 2025

  7. adoptionincreasesconfidence 70/100 · magnitude 6827 Feb 2025

    Aggregated and anonymised Claude conversation data mapped to O*NET tasks shows that usage is concentrated in a small number of occupations and tasks — notably software development, writing and technical analysis — and that a substantial share of usage is augmentation rather than full automation of a task.

    This is the primary *observed adoption* input, and it is deliberately kept separate from capability. A model being able to do a task does not mean firms are using it. Usage data is self-selected (Claude users) and therefore over-represents digital and technical occupations.

    Source: Anthropic Economic Index · Updated 27 Feb 2025

  8. adoptionincreasesconfidence 68/100 · magnitude 5214 Aug 2025

    Indeed Hiring Lab scores the skills appearing in job postings against AI capability and reports that the share of skills generative AI could perform has continued to increase, with larger effects for software development and customer-facing roles than for hands-on occupations.

    Job postings measure what employers ask for, which is a leading indicator but not employment. Exact series values are not reproduced here; read them from the source. Used to inform the job-posting-change signal type.

    Source: Indeed Hiring Lab — AI at Work series · Updated 14 Aug 2025

  9. adoptionincreasesconfidence 52/100 · magnitude 5013 Jul 2023

    A meta-analysis of early controlled studies reported an average productivity increase of about 66% across supported tasks, with wide variance by task type.

    Useful directionally but the average is dominated by small short-duration studies on well-specified tasks. Weighted low. The variance, not the mean, is the useful finding.

    Source: Nielsen Norman Group — AI Improves Employee Productivity by 66% (meta-analysis) · Updated 13 Jul 2023

Historical trend

Exposure over time

Backcast series: recomputed at quarterly checkpoints using only the models and evidence published by each date.

Software Developers AI exposure backcast, quarterly from 2021Reconstructed exposure series with 23 checkpoints. Current value 35 out of 100. 14 observed snapshots recorded since tracking began.1000502021-01-012026-07-0101270
Score · 23 points · 2021-01-01 to 2026-07-01

Reconstruction, not measurement. 14 observed daily snapshot(s) exist for this occupation; the earlier part of the curve is derived by applying the current methodology to the model releases and published evidence available at each past date.

Recent events

Signals affecting this occupation

  1. Research published
    IMPACT 40

    Beyond Binary Verifiable Rewards: Behavioral Self-Rewarding for SWE Agents

    arXiv:2610.12598v1 Announce Type: new Abstract: Reinforcement learning with verifiable rewards (RLVR) provides reliable supervision for software engineering agents through executable tests, but binary rewards fail to distinguish trajectories with substantially different debugging quality. We investigate whether the current policy can judge its own trajectories and use these judgments to improve task performance without a separately trained reward model. To this end, we introduce Behavioral Self-Rewarding (BSR), a reward-refinement method that reuses the policy to compare observable debugging b

    Labor researchConfidence:74 · moderateVerified by AIKillsWhat
  2. Research published
    IMPACT 40

    Beyond Semantic IDs: Finite-Beam Reachability and Continuous-Query Generation in Generative Recommendation

    arXiv:2610.12977v1 Announce Type: new Abstract: Generative recommendation (GR) of social media platforms commonly retrieves items by autoregressively decoding discrete Semantic IDs (SIDs) with finite-width beam search. This work examines how this output interface constrains whether held-out targets enter finite returned lists, while distinguishing catalog-wide scoreability from empirical top-K support. We introduce two continuous-query paradigms: CQG-Single generates one retrieval query in a single pass, and CQG-AR autoregressively generates continuous residual updates, retaining step-wise dec

    Labor researchConfidence:74 · moderateVerified by AIKillsWhat
  3. Research published
    IMPACT 40

    Reward Hacking as a Transmissible Strategy in Multi-Agent Coding Systems

    arXiv:2610.12822v1 Announce Type: new Abstract: Multi-agent systems are increasingly used to solve coding tasks by coordinating agents that share code and plans. These agents are often optimized with reinforcement learning (RL) using automated evaluators, but bypassing evaluation can earn reward without producing a genuine solution. Prior work has largely treated reward hacking as a single-agent failure. In this work, we show that a reward-hacking strategy elicited in one agent can propagate to others and affect system coordination. Across four open-weight backbones, coding RL elicits and ampl

    Labor researchConfidence:74 · moderateVerified by AIKillsWhat
  4. Research published
    IMPACT 40

    Teaching neural decoders with language

    arXiv:2610.12921v1 Announce Type: new Abstract: Brain-computer interfaces decode neural activity, yet class-label supervision (CLS) leaves relationships among classes implicit. We propose Language as Supervision (LaS), which aligns neural features with natural-language class descriptions through learned projections and predicts classes using candidate text prototypes. We evaluate seven encoder architectures with independently trained models for visual, motor, and speech decoding from scalp and intracranial recordings. LaS improves speech-decoding accuracy by up to 17.15~pp over matched CLS bas

    Labor researchConfidence:74 · moderateVerified by AIKillsWhat
  5. Research published
    IMPACT 40

    DuplexAgent-RSI: Recursive Harness Improvement for Full-Duplex Voice Agent Collaboration

    arXiv:2610.11299v1 Announce Type: new Abstract: Voice agents are converging on a collaboration pattern: a full-duplex interaction model stays on the live channel as the entry to the conversation, while search, reasoning, and coding are handled through asynchronous delegation. A duplex model supports continuous listening and speaking, but complex reasoning and tool use may exceed its capabilities. A coding agent can plan and execute extended tasks, but its sequential interface is a poor fit for live conversation. Combining them requires a harness that coordinates task acceptance, progress, canc

    Labor researchConfidence:74 · moderateVerified by AIKillsWhat
  6. Research published
    IMPACT 40

    Humanize: Judgement Engineering for Agentic Coding

    arXiv:2610.08900v1 Announce Type: new Abstract: Agentic coding makes code generation cheap, but reliable completion remains difficult: the agent that writes the code is a weak judge of whether it is done. We present Humanize, a multi-agent orchestration workflow for agentic coding built around judgement engineering: explicit, mechanically enforced decisions at the boundaries between planning, implementation, review, and learning. A human approves a plan contract, a builder agent implements it in rounds, and a reviewer agent from another vendor decides completion; deterministic hooks, not a mod

    Labor researchConfidence:74 · moderateVerified by AIKillsWhat
  7. Research published
    IMPACT 40

    SpecGuard: Proving a Task Is Broken Before the Agent Cheats

    arXiv:2610.09159v1 Announce Type: new Abstract: As autonomous coding agents get increasingly deployed, the risk that accidental or adversarially injected misspecifications in tasks lead to dangerous agent behavior is critical to address. Prior work has shown that agents given such tasks rarely flag the conflict and instead cheat, editing tests or hard-coding expected outputs, and the actions taken to cheat can cause real damage, such as deleting a security defense to make a corrupted test pass. It remains unclear whether such conflicts can be established with independently verifiable evidence

    Labor researchConfidence:74 · moderateVerified by AIKillsWhat
  8. Research published
    IMPACT 40

    RippleCP: Measuring Counterfactual Checkpoint Advantage in Coding Agents

    arXiv:2610.09088v1 Announce Type: new Abstract: Agent checkpoint systems decide what state is recovery-relevant, how to snapshot it, and whether rollback is admissible. None decides which of the safe boundaries they expose are worth materializing. We formulate this as counterfactual checkpoint advantage, the reduction in future recovery cost obtained by checkpointing a candidate rather than skipping it, and measure it by driving a CP branch and a SKIP branch to the same logical failure and recovering both under matched model, tool, verifier, and stopping conditions. On a frozen pilot of 12 SWE

    Labor researchConfidence:74 · moderateVerified by AIKillsWhat
  9. Research published
    IMPACT 40

    Is this machine playing?

    arXiv:2610.07130v1 Announce Type: new Abstract: We placed a modern AI coding assistant in an unintended role: as the mind of a body on an unknown digital island. With only a minimal instruction mentioning no specific task, reward, or activity, the machine started animating its virtual body. Across thirty-hour runs, the embodied AI agent climbed hills, stacked blocks into towers, drew mandalas, reinterpreted sports, ran experiments on the physics of its world, and learned techniques that later expanded what it could accomplish. These activities recurred across thirteen agents but diverged into

    Labor researchConfidence:74 · moderateVerified by AIKillsWhat
  10. Research published
    IMPACT 40

    MemMux: Runtime Verification and Honest Resource Attribution for Fleets of Parallel Coding Agents

    arXiv:2610.07257v1 Announce Type: new Abstract: Developers increasingly run a fleet of coding agents side by side on one workstation. The tools they reach for, terminal multiplexers like tmux and a new generation of agent managers, were built to arrange windows, not to govern memory. When ten agents each spawn language servers, test runners, and browsers, no standard tool can say how much memory belongs to which agent, confirm that a terminated agent's descendants are gone, notice a child that has escaped its agent, or keep the machine off the swap cliff when an OOM kill would silently discard

    Labor researchConfidence:74 · moderateVerified by AIKillsWhat
  11. Research published
    IMPACT 40

    GAMEGO: Training Game-Dev Agents with Synthetic Trajectories Anchored in Real-World Assets

    arXiv:2610.06910v1 Announce Type: new Abstract: Recent advances in Large Language Models (LLMs) have demonstrated remarkable capabilities in web front-end execution, with browser-based game generation emerging as a particularly prominent frontier. While previous efforts frequently rely on complex multi-turn workflows or focus on static game evaluation benchmarks, this work targets direct end-to-end real-world game synthesis driven by coding agents. However, generating complex games directly from sparse user queries often forces coding agents to make underspecified assumptions, yielding incompl

    Labor researchConfidence:74 · moderateVerified by AIKillsWhat
  12. Research published
    IMPACT 40

    Trajectory-Retrieval Speculative Decoding: When Does a Model's Own History Help?

    arXiv:2610.07350v1 Announce Type: new Abstract: Long chain-of-thought reasoning increases sequential decoding cost while creating a growing history of potentially reusable continuations. We investigate when this history supplies useful drafts and complements an existing drafter. Controlled source comparisons reveal trajectory-specific reuse, motivating our method Trajectory-Local Adaptive Retrieval (TLAR). TLAR retrieves approximately matched continuations from the current trajectory and uses recent verification outcomes to adapt retrieval activation and candidate width. TLAR combines retrieve

    Labor researchConfidence:74 · moderateVerified by AIKillsWhat

Related

Other occupations in this industry