Give agents whole tasks before giving them whole jobs
EM outlook · –
During 2027–2028, we expect delegation to expand fastest where outcomes are easy to check and failures are recoverable. Rising software-task horizons make that path credible. Broad autonomy faces a tougher test: completing a task half the time is far from dependable service, and already finds human attention under pressure.
What this means
Move from checking each step to checking outcomes where you can detect a failure and recover. For harder work, preserve checkpoints until actual completion records justify removing them.
Why we take this view
This predicts wider dependable delegation of bounded tasks over the next two years. Clear success tests and recovery can turn existing capabilities into a usable service sooner than broad, unsupervised responsibility.
What has to hold
The task has a clear success condition and failures can be detected before their cost becomes unacceptable.
Verification and recovery improve along with model capability.
The strongest rival
Longer tasks multiply hidden failures; humans remain closely involved even as agents produce more.
Agent systems learn to diagnose and recover across workflows, allowing whole-job responsibility to grow faster than this forecast.
What would change our mind
In workflows selected for clear outcomes and recoverable failures, repeated deployments still require frequent rescue or show persistent hidden defects despite better scores.
Higher success thresholds improve alongside task length; gains survive when tasks are unfamiliar and consequences are real.
Measure completion, hidden failure, human intervention and recovery in real deployments as well as .
Where this could lead
Whole tasks move off the human’s desk
Base case · –
Agents complete specified work with outcome checks and exception handling. People spend less time approving steps, while retaining responsibility for goals and consequential decisions.
What would make it happen
Verification and recovery around bounded tasks can make existing and improving capabilities usable before broad autonomous judgment arrives.
Conditions
Success can be tested and a failed run can be retried, reversed or escalated affordably.
Context and permissions match the task instead of requiring continual human clarification.
Watch for
The share of accepted tasks completed without rescue rises at a stable quality bar.
Failure and recovery costs fall across repeated real deployments, not just runs.
Capability improves, but long tasks accumulate mistakes and missing context. Businesses keep people close enough to catch problems before they become expensive.
What would make it happen
Over the same adoption period, reliability can lag capability where outcomes are ambiguous or recovery is difficult.
Conditions
High-consequence failures remain difficult to detect before acting.
Independent task success does not transfer reliably across unfamiliar conditions.
Watch for
Task length improves at moderate success rates while dependable completion lags.
Deployments continue to need frequent interventions or hide substantial human rescue.
Why this connects to When can work be delegated without checking every step?
EM interpretation
A result can be useful because a person caught its mistakes. Removing that review changes the job: the agent must detect failure, recover or ask for help. That is why productivity gains can arrive before dependable autonomy.
Track the human interventions behind a successful result before applying it to unsupervised work.
Joel Becker, Nate Rush, Tom Cunningham, David Rein and Khalid Mahamud
authors of the follow-up methods update · 2026-02-24 · From the author’s own work
judges that developers likely benefited more from AI in early 2026 than in early 2025, while warning that participation, task selection and time measurement make the size of that improvement hard to estimate.
Sources and passages
Publication date of the authored source.
This is a later team’s assessment of a changed experiment, not the same measurement repeated cleanly.
chief executive writing a personal essay · 2024-10 · From the author’s own work
In October 2024, Amodei suggested that powerful AI could arrive as early as 2026 while explicitly allowing a much longer wait. He defined it through exceptional breadth and autonomous work, then explored possible consequences under that assumption.
Sources and passages
Month printed in the essay; the underlying capture has no publication timestamp.
The essay’s definition is not a universally agreed endpoint.
An earliest possibility is not a median date, deadline or claim that the system exists.
Publisher publication date, checked on original page.
Original publisher page and exact retained source reviewed. Source retention does not establish every downstream reuse or training right.
Exact retained Stream work/snapshot and SHA256 verified against the 29 October SELECT-only source packet. That lookup found no joined public Stream or bridge reading revision; no reading identity is invented.
This acceleration makes human attention the bottleneck for agent output.
Retained extracted text, character offsets [767, 839). Context (EM paraphrase): The authors report rising internal output alongside agent use and interpret human attention as a constraint. Adoption is not randomized; the report does not establish a general causal productivity effect.
Publication date displayed by the original publisher.
Original publisher page and exact retained source reviewed. Source retention does not establish every downstream reuse or training right.
Exact retained Stream work/snapshot and SHA256 verified against the 29 October SELECT-only source packet. That lookup found no joined public Stream or bridge reading revision; no reading identity is invented.
16 experienced developers
Retained extracted text, character offsets [1527, 1552). Context (EM paraphrase): Methods: developers worked on 246 real issues in familiar open-source repositories using early-2025 tools. This population does not represent all developers or later systems.
they take 19% longer to complete issues
Retained extracted text, character offsets [2468, 2507). Context (EM paraphrase): Core Result, first paragraph; interpret with Methods and the immediately following limitations table. Reported result for the 2025 study, not a current universal estimate.
Publication date displayed by the original publisher.
Original publisher page and exact retained source reviewed. Source retention does not establish every downstream reuse or training right.
Exact retained Stream work/snapshot and SHA256 verified against the 29 October SELECT-only source packet. That lookup found no joined public Stream or bridge reading revision; no reading identity is invented.
the data from our new experiment gives us an unreliable signal of the current productivity effect of AI tools
Retained extracted text, character offsets [431, 540). Context (EM paraphrase): Opening discussion, paragraph beginning Unfortunately. Subsequent paragraphs describe participation, task selection and time measurement.
likely biases downwards
Retained extracted text, character offsets [715, 738). Context (EM paraphrase): METR says nonparticipation by developers unwilling to work without AI likely understates speedup; the authors believe early-2026 tools speed developers more than early-2025 tools. Lower study pay, task selection and concurrent-agent time accounting limit the size estimate.
October 2024 is printed in the retained essay. The work metadata has no publication timestamp; month precision is preserved.
Original personal essay retained in Stream and reviewed here; not an institutional position or a new 2026 statement. Public availability does not establish every reuse right.
Exact original Stream identity and SHA256 verified against the retained source packet.
as early as 2026
Retained extracted text, character offsets [6255, 6271). Context (EM paraphrase): Amodei says powerful AI could arrive that early but could also take much longer. His working definition combines exceptional cognitive capability, human-like computer interfaces and autonomous work over hours, days or weeks. The essay assumes relatively near arrival to explore later effects; it is not a demonstrated outcome or a standardized AGI forecast.
Original methodology update checked. Read alongside the September dashboard status; January is not presented as the latest frontier measurement.
Original public source read on 30 October 2026. An exact retained Stream identity has not been resolved for this edition.
a steady exponential increase
Opening account of the earlier time-horizon work. EM context: the January update reports continued growth through 2025, but growth estimates depend on task composition; this is historical software-task evidence, not a current AGI clock.
Page last-updated date, not its original publication date.
labels this page no longer actively updated, last updated 8 September 2026. It supports a defined and its limits, not a complete October model comparison.
Original public source read on 30 October 2026. An exact retained Stream identity has not been resolved for this edition.
predicted to succeed half the time
Opening definition of the 50%-time horizon. EM context: duration means human expert task time, not agent runtime. Tasks are primarily software engineering, machine learning and cybersecurity, with explicit success criteria.
Measurements above 16 hrs are unreliable with our current task suite
Notice above the interactive task plot, checked 30 October 2026.
Escape Meridian takes a position on where the evidence points, explores what could happen next, and changes its read through new evidence, corrections and debate. The outlooks are our judgments; the sources retain their own voices.
About this edition
EM outlooks authored 30 October 2026 from the retained 29 October source review, supplemented by five original public sources checked on 30 October. Employment evidence centers on US workers and selected workplaces; capability evidence centers on software tasks. Outlook dates express EM judgments about the stated outcomes, not dates forecast by the cited authors. Source publication, access and evidence-review dates remain distinct.
