# From tab autocomplete to runs you sleep through · Oxagen Research

Source: https://oxagen.sh/blog/from-tab-autocomplete-to-runs-you-sleep-through

[Research](https://oxagen.sh/blog) · [Autonomous agents](https://oxagen.sh/blog/pillars/autonomous-agents)

# From tab autocomplete to runs you sleep through

Agent-assisted coding grew from a completion shortcut to multi-day autonomous runs in about five years. This post traces the measurements at each step.

Oxagen Research · November 10, 2026 · 5 min read

*   [Autonomous agents](https://oxagen.sh/blog/pillars/autonomous-agents)
*   [AI agents](https://oxagen.sh/blog/pillars/ai-agents)

In 2021, AI assistance in an editor meant pressing tab. The model finished the line you had already started, and you accepted or ignored it one line at a time. The median accepted suggestion was short, the failure mode was a wrong identifier, and the cost of being wrong was one undo. A developer could supervise every token the system produced, because there were few tokens and they arrived while the developer was already looking at the screen.

Five years later, teams hand an agent a ticket on Friday and read a pull request on Monday. The jump from one to the other is the largest change in how software gets written since the IDE, and it happened faster than the process around it could follow. Each step along the way looked small. The accumulated distance was not.

## The measurements at each step

The progression has public numbers at most of its points, which is rare for a tooling shift this fast.

Five years of agent-assisted coding, by the unit of work

| Fact | Holds | Status (Now) |
| --- | --- | --- |
| Tab autocomplete: the unit is a line | 2021 to 2023 | No longer holds |
| Chat in the editor: the unit is a file | 2023 to 2025 | No longer holds |
| Agent runs: the unit is a passing task | 2025 onward | Holds |
| Measured horizon doubles each ~7 months | 2019 onward | Holds |

Each row is a tool that shipped and was measured, not a forecast. The bottom row is METR's fitted doubling; the other three are the form the tool took at the time.

The first step was autocomplete, and the measurements were modest. GitHub reported that the 1.0 release of GitHub Copilot, in mid-2022, generated about 27 percent of the code in files where it was enabled, rising to a majority in Java.[1](https://oxagen.sh/blog/from-tab-autocomplete-to-runs-you-sleep-through#user-content-fn-1) The controlled studies that followed found speed-ups on bounded tasks: one trial with 95 professional developers found the group given GitHub Copilot finished an HTTP-server task 55.8 percent faster.[2](https://oxagen.sh/blog/from-tab-autocomplete-to-runs-you-sleep-through#user-content-fn-2) The tasks were short and self-contained. That detail mattered, and it went unquoted.

The second step was chat. The unit of assistance grew from a line to a function to a file, but the loop stayed the same: a person wrote the prompt, read the output, pasted what survived, and ran the tests. The person was still in every round trip.

The third step removed the person from the loop's middle. Agentic coding tools, such as Claude Code, Codex, and their peers, took over the read-edit-run cycle entirely. The agent reads the failure, edits the code, runs the test again, and repeats until the task passes or the budget ends. The person moved to the two ends: stating the task and reviewing the result.

## Adoption arrived before the process did

The survey numbers describe a faster uptake than any prior developer tool. JetBrains' 2024 survey found 61 percent of developers already using AI tools regularly for code-related work.[3](https://oxagen.sh/blog/from-tab-autocomplete-to-runs-you-sleep-through#user-content-fn-3) Stack Overflow's 2024 survey put the share at 62 percent using or planning to use, and its 2025 follow-up put the daily-use share above half.[4](https://oxagen.sh/blog/from-tab-autocomplete-to-runs-you-sleep-through#user-content-fn-4) DORA's 2024 report found that more than 75 percent of respondents relied on AI for at least one daily professional responsibility.[5](https://oxagen.sh/blog/from-tab-autocomplete-to-runs-you-sleep-through#user-content-fn-5)

The same surveys carry the warning. Stack Overflow's 2025 data shows trust falling as usage rose: more respondents distrusted the accuracy of AI output than trusted it, and the gap widened among the heaviest users. DORA found a 25 percent increase in AI adoption linked to an estimated 1.5 percent drop in delivery throughput and a 7.2 percent drop in delivery stability, while about three quarters of the same respondents reported feeling more productive.[5](https://oxagen.sh/blog/from-tab-autocomplete-to-runs-you-sleep-through#user-content-fn-5) The tool got adopted faster than the workflow around it got rebuilt. We covered where that cost lands in _The problem is not slop, it is your process_.

## How run length became the number that matters

The cleanest measure of the whole arc is METR's task-completion time horizon: how long a task an agent can finish on its own, measured by the time a skilled human would need. When the work appeared in March 2025, the frontier 50 percent horizon was about 50 minutes, doubling roughly every seven months since 2019.[6](https://oxagen.sh/blog/from-tab-autocomplete-to-runs-you-sleep-through#user-content-fn-6) The January 2026 revision put the doubling time at 196.5 days overall and 130.8 days for models released since 2023, with the highest measured horizon at 320 minutes.[7](https://oxagen.sh/blog/from-tab-autocomplete-to-runs-you-sleep-through#user-content-fn-7) We worked through the arithmetic in _The agent time horizon is doubling_.

Those numbers compress the five-year story into one sentence. A tool that once finished your line now finishes tasks that take a person hours, and the length it handles is doubling on a schedule measured in months. METR states its own limit plainly: measurements above 16 hours are unreliable with the current task suite, because no suite can grade a week-long run.[8](https://oxagen.sh/blog/from-tab-autocomplete-to-runs-you-sleep-through#user-content-fn-8) The trend line and the measurement apparatus are racing, and the trend line is ahead.

## What long runs actually break

A multi-day run does not fail the way a long autocomplete suggestion fails. The failure mode changed with the unit of work.

An autocomplete error is visible in the token where it happens. A long run's error is a wrong turn that becomes load-bearing: the agent writes the mistake into its own context, then spends hours staying consistent with it. Sinha and colleagues measured this as self-conditioning: errors already in the model's context raise the rate of later errors.[9](https://oxagen.sh/blog/from-tab-autocomplete-to-runs-you-sleep-through#user-content-fn-9) Backlund and Petersson's Vending-Bench found the same shape at scale: agents running a simulated business fell into loops they rarely recovered from, with no clear link to the context window filling.[10](https://oxagen.sh/blog/from-tab-autocomplete-to-runs-you-sleep-through#user-content-fn-10) The run works correctly, takes one wrong turn, and then builds forty consistent steps on top of it.

This is why the supervision question changed. Watching worked when the unit was a line. It barely worked when the unit was a function. It does not work when the unit is a week of work, because no person reads that transcript, and the step that mattered is indistinguishable from the ones that did not until someone queries for it.

## How Oxagen handles this

Oxagen is workforce management for autonomous agents. It does not run them. It holds each agent's mandate: the identity it acts as, the systems and data it may request, the budget and rules it works under, the tools and skills it is equipped with, and the record of what it did.

The arc this post describes maps onto three of those clauses. As runs lengthened, the decisions a reviewer used to make mid-run moved into the mandate: what the agent may reach, what it may spend, and which requests stop and wait for a named person. Oxagen answers each governed request at the moment of use, so a rule set before the run still applies at hour forty. It records the cost of each governed action, assigned to the person, the agent, the run, the turn, and the step. And it keeps the record as a series of frames next to the mandate, so finding the wrong turn in a week-long run is a query, not a reread.

The tooling went from tab to a working week in five years. The mandate is what a team has when the next doubling arrives.

## References

1.  GitHub (2022). _Research: Quantifying GitHub Copilot's impact on developer productivity and happiness_. [https://github.blog/news-insights/research/research-quantifying-github-copilots-impact-on-developer-productivity-and-happiness/](https://github.blog/news-insights/research/research-quantifying-github-copilots-impact-on-developer-productivity-and-happiness/) [↩](https://oxagen.sh/blog/from-tab-autocomplete-to-runs-you-sleep-through#user-content-fnref-1)
    
2.  Peng, S., Kalliamvakou, E., Cihon, P., & Demirer, M. (2023). _The Impact of AI on Developer Productivity: Evidence from GitHub Copilot_. arXiv. [https://arxiv.org/abs/2302.06590](https://arxiv.org/abs/2302.06590) [↩](https://oxagen.sh/blog/from-tab-autocomplete-to-runs-you-sleep-through#user-content-fnref-2)
    
3.  JetBrains (2024). _The State of Developer Ecosystem 2024_. [https://www.jetbrains.com/lp/devecosystem-2024/](https://www.jetbrains.com/lp/devecosystem-2024/) [↩](https://oxagen.sh/blog/from-tab-autocomplete-to-runs-you-sleep-through#user-content-fnref-3)
    
4.  Stack Overflow (2024, 2025). _Developer Survey_. [https://survey.stackoverflow.co/](https://survey.stackoverflow.co/) [↩](https://oxagen.sh/blog/from-tab-autocomplete-to-runs-you-sleep-through#user-content-fnref-4)
    
5.  DORA (2024). _Accelerate State of DevOps Report 2024_. Google Cloud. [https://dora.dev/research/2024/dora-report/](https://dora.dev/research/2024/dora-report/) [↩](https://oxagen.sh/blog/from-tab-autocomplete-to-runs-you-sleep-through#user-content-fnref-5) [↩2](https://oxagen.sh/blog/from-tab-autocomplete-to-runs-you-sleep-through#user-content-fnref-5-2)
    
6.  Kwa, T., West, B., Becker, J., et al. (2025). _Measuring AI Ability to Complete Long Tasks_. arXiv. [https://arxiv.org/abs/2503.14499](https://arxiv.org/abs/2503.14499) [↩](https://oxagen.sh/blog/from-tab-autocomplete-to-runs-you-sleep-through#user-content-fnref-6)
    
7.  METR (2026). _Time Horizon 1.1_. [https://metr.org/blog/2026-1-29-time-horizon-1-1/](https://metr.org/blog/2026-1-29-time-horizon-1-1/) [↩](https://oxagen.sh/blog/from-tab-autocomplete-to-runs-you-sleep-through#user-content-fnref-7)
    
8.  METR. _Task-Completion Time Horizons of Frontier AI Models_. [https://metr.org/time-horizons/](https://metr.org/time-horizons/) [↩](https://oxagen.sh/blog/from-tab-autocomplete-to-runs-you-sleep-through#user-content-fnref-8)
    
9.  Sinha, A., Arun, A., Goel, S., Staab, S., & Geiping, J. (2025). _The Illusion of Diminishing Returns: Measuring Long Horizon Execution in LLMs_. arXiv. [https://arxiv.org/abs/2509.09677](https://arxiv.org/abs/2509.09677) [↩](https://oxagen.sh/blog/from-tab-autocomplete-to-runs-you-sleep-through#user-content-fnref-9)
    
10.  Backlund, A., & Petersson, L. (2025). _Vending-Bench: A Benchmark for Long-Term Coherence of Autonomous Agents_. arXiv. [https://arxiv.org/abs/2502.15840](https://arxiv.org/abs/2502.15840) [↩](https://oxagen.sh/blog/from-tab-autocomplete-to-runs-you-sleep-through#user-content-fnref-10)