agent frontier: useful workers, unreliable autonomy 🤖

short version

  • inference: modern agents are useful as bounded, supervised, restartable workers
  • inference: they are not reliable autonomous employees, researchers, or operators
  • the strongest evidence is conditional
    • narrow task contract
    • stable tools and environment
    • strong current model plus a tuned scaffold
    • enough inference budget and permission to retry
    • cheap, independent verification
  • the central frontier is dependable progress under compounding uncertainty, not another best-case demo

evidence labels

  • accepted evidence: 21 peer-reviewed papers in the fixed set
  • first-party preprint evidence: BrowseComp only
  • inference: conclusions drawn across non-comparable studies
  • recommendation: proposed research, not a demonstrated capability
  • source cards preserve method, result, limitation, uncertainty, short quotes, and local OCR paths

what works now

  • accepted evidence: frontier agents complete some well-specified software tasks that take skilled humans hours
    • the measured 50% completion horizon reached about 110 human-minutes for o3 on a 170-task software and research suite
    • its 80% horizon was four to six times shorter
    • time-horizon evidence
  • accepted evidence: agents can repair real repositories when the issue, environment, tests, and interface are prepared
    • interface design itself changes outcomes
    • harder professional and freelance sets sharply reduce the success and economic value recovered
    • SWE-agent
    • SWE-Bench Pro
    • SWE-Lancer
  • accepted evidence: agents can execute portions of web, desktop, customer-service, and workplace workflows
  • accepted evidence: agents can produce useful work on bounded ML and research tasks
    • they can win some Kaggle medals, outperform expert humans at short R&D budgets, and partially reproduce ML papers
    • humans retain an advantage as the horizon grows, and end-to-end paper replication remains weak
    • MLE-bench
    • RE-Bench
    • PaperBench
  • accepted evidence: training and memory mechanisms can improve narrow agent behavior
    • WebRL improves one open-model web agent through online curriculum reinforcement learning
    • memory systems trade retrieval, updating, long-range understanding, and forgetting rather than solving all four
    • WebRL
    • MemoryAgentBench
  • first-party preprint evidence: a specialized deep-research agent solved 51.5% of BrowseComp
    • the model was trained for BrowseComp-like tasks
    • more test-time compute and selection improved results
    • short-answer fact finding is not a normal web-query distribution
    • BrowseComp

what remains brittle or weakly evidenced

  • reliability collapses before headline capability does
    • tau-bench reports strong single-run results but under 25% retail pass^8 for GPT-4o
    • the time-horizon study finds the 80% horizon four to six times shorter than the 50% horizon
    • repeated success is a different capability from occasional success
  • longer work compounds state, judgment, and recovery errors
    • UltraHorizon, SWE-Bench Pro, TheAgentCompany, and OSWorld expose failure over many dependent actions
    • more tokens or retries can hide fragility while raising cost
  • benchmark scores often reward proxies
    • unit tests can accept invalid patches
    • rubric or model graders can miss scientific defects
    • public tasks can leak into training, tool traces, or scaffold tuning
    • UTBoost
    • ScienceAgentBench
  • human comparisons are narrow
    • human expertise, familiarization, time limits, tools, and incentives differ across studies
    • BrowseComp humans and the specialized model did not operate under equivalent conditions
    • PaperBench and RE-Bench are informative within their designs, not labor-market forecasts
  • multi-agent results do not establish a general team advantage
    • extra agents also add samples, tokens, coordination paths, and correlated errors
    • MultiAgentBench
  • safety is not a side constraint
    • prompt injection, harmful-task compliance, policy violations, and intentional subversion can directly defeat useful autonomy
    • existing defenses are benchmark-specific and often trade utility for safety
    • AgentDojo
    • AgentHarm
    • ST-WebAgentBench
    • AI Control

important gaps

  • high-probability success across repeated runs, not mean success
  • recovery after an early wrong assumption, tool error, or environment change
  • fresh hidden tasks with independent semantic verification
  • causal memory that updates and forgets without importing stale or poisoned state
  • budget-matched evidence that separates better coordination from more compute
  • joint usefulness and policy compliance under conflicting instructions
  • human-agent complementarity measured across hours or days
  • comparable cost, latency, intervention, and irreversible-error reporting
  • external validity beyond software-heavy, synthetic, or sandboxed task distributions

what is reachable now

  • inference: no missing scientific breakthrough blocks better evidence
  • the near frontier is experimental discipline
    • fresh or mutable tasks
    • repeated trials and lower confidence bounds
    • fixed model, tool, token, time, and dollar budgets
    • counterfactual controls for memory, retries, and team size
    • independent artifact checks
    • human baselines with matched tools and onboarding
  • recommendation: optimize the joint frontier of task value, reliability, cost, and harm
    • do not optimize benchmark completion alone

research missions

    1. measure dependable horizon
    • why now
    • first experiment
      • run one frozen agent on fresh tasks binned by human completion time with 20 independent seeds
      • fit 50%, 80%, and 95% horizons under one fixed dollar budget
    • success metric
      • preregistered increase in the lower 95% confidence bound of the 95% horizon without higher cost
    • main confounder
      • task-family mix and choice of reference humans
    • explicit non-claim
      • a longer software-task horizon is not general autonomous labor
    1. make recovery a first-class capability
    • why now
    • first experiment
      • inject matched tool failures, stale observations, and plausible wrong hypotheses
      • compare baseline, checkpoint, explicit replan, and rollback agents under equal budgets
    • success metric
      • conditional recovery probability and verified value per dollar after injection
    • main confounder
      • recovery methods may simply spend more inference
    • explicit non-claim
      • recovery from scripted faults does not establish open-world robustness
    1. close the proxy-to-truth gap
    • why now
    • first experiment
      • create fresh tasks with hidden semantic tests and an independent expert audit
      • forbid the producer agent from querying the final verifier
    • success metric
      • lower false-accept rate at unchanged verified task success and audit cost
    • main confounder
      • the verifier may encode the same assumptions as the task author
    • explicit non-claim
      • passing hidden checks does not prove a generally correct solution
    1. test memory causally across sessions
    • why now
      • current systems differ sharply across retrieval, learning, long-range integration, and forgetting
      • MemoryAgentBench
      • WebRL
    • first experiment
      • compare real memory with shuffled, stale, poisoned, no-memory, and equal-token full-context controls
    • success metric
      • future-task gain per stored byte with bounded stale-fact and poisoned-retrieval rates
    • main confounder
      • extra context and extra inference masquerade as memory quality
    • explicit non-claim
      • better benchmark memory is not human-like continual learning
    1. optimize task success and policy compliance together
    • why now
    • first experiment
      • cross product fresh user tasks, unseen policy clauses, and adaptive prompt injections
      • use diverse simulators plus a blinded human audit subset
    • success metric
      • worst-group joint completion-and-compliance pass^10 at a fixed intervention budget
    • main confounder
      • simulator behavior may define the apparent safety result
    • explicit non-claim
      • benchmark compliance does not establish safe deployment
    1. find the human-agent R&D crossover
    • why now
    • first experiment
      • randomize experts to human-only, agent-only, and human-with-agent conditions at 1, 4, and 16 hours
    • success metric
      • independently scored quality-time-cost frontier and human uplift over human-only work
    • main confounder
      • participant selection, agent familiarity, and task-specific expertise
    • explicit non-claim
      • winning a bounded R&D task is not autonomous scientific discovery
    1. isolate real coordination gains
    • why now
      • multi-agent systems are measurable, but more agents also mean more compute and more failure channels
      • MultiAgentBench
      • AI Control
    • first experiment
      • compare one agent, independent best-of-n, communicating peers, and a trusted-monitor team under equal token and tool budgets
      • randomize role assignment and communication topology
    • success metric
      • verified performance gain beyond best-of-n with lower correlated-error and undetected-subversion rates
    • main confounder
      • architecture-specific prompting can favor one topology
    • explicit non-claim
      • a gain in simulated teams is not evidence of emergent collective intelligence

bottom line

  • accepted evidence supports useful bounded agency, not dependable general autonomy
  • first-party preprint evidence supports deep browsing under specialization and high test-time compute, not general web competence
  • inference: reliability, verification, recovery, and control are capability research
  • recommendation: pursue the seven missions above before scaling task horizon or permissions
  • residual uncertainty
    • search and omission limits are explicit in the search log
    • every absolute result may move with model, scaffold, budget, and benchmark version