85 agentic tools audited for security incidents · Talk to the lab →
MotionticBook a call
Glossary

What is deployment overhang?

Deployment overhang is the gap between the autonomy an AI model can handle and the autonomy it is actually given in real deployments, where people and products grant less latitude than the model could manage.

The term appears in Anthropic's February 2026 research post, Measuring AI agent autonomy in practice. The post compares what models can do in controlled tests with how long and how independently they actually run for real users, and finds that the latitude granted in practice lags behind what the models can handle.

The overhang is not an argument for giving agents more freedom straight away. It describes where the bottleneck sits: in many deployments, the limit on what an agent does is product design, oversight and trust rather than model capability.

The evidence

  • What models can do

    METR estimates that Claude Opus 4.5 completes, with 50% success, tasks that take a person nearly 5 hours. Anthropic notes this measures task difficulty in an idealised setting, not how long a model runs, so the figures are not directly comparable.

  • How long they actually run

    The median Claude Code turn lasts about 45 seconds. The 99.9th percentile turn nearly doubled, from under 25 minutes to over 45, between October 2025 and January 2026, smoothly across model releases.

  • Internal use

    From August to December, success on Anthropic's internal users' hardest tasks doubled while human interventions per session fell from 5.4 to 3.3.

  • Few irreversible actions

    Across nearly a million public API tool calls, 0.8% appeared irreversible, and 80% came from agents with at least one safeguard.

How oversight changes as trust grows

New Claude Code users, with under 50 sessions, use full auto-approve in about 20% of sessions; by 750 sessions the share passes 40%. Experienced users also interrupt more often, in about 9% of turns against about 5% for new users. Anthropic reads this as oversight moving from approving each step to monitoring and stepping in when something looks wrong.

The model takes part too. On the most complex tasks, Claude stops to ask for clarification more than twice as often as on minimal ones, and it stops to ask more often than people interrupt it.

What Anthropic recommends

  • Post-deployment monitoring

    Pre-deployment evals cannot show how people actually use agents, so model and product developers should invest in monitoring after launch.

  • Models that flag uncertainty

    Train models to recognise their own uncertainty and raise issues with a person.

  • Products built for oversight

    Make agent actions visible and intervention simple.

  • No mandated approval patterns

    Policymakers should avoid requiring approval of every action, which adds friction without clear safety benefit, and focus on whether people can monitor and intervene.

What it does not mean

A general overhang does not mean a given agent is ready for more autonomy. The figures are aggregates dominated by software work, which made up nearly half of tool calls in the API sample, and the capability estimate comes from a different measurement than the usage data. A specific agent earns autonomy through its own eval results and production record, one tool and one risk tier at a time.

How Motiontic's engineers apply it

Motiontic's engineers treat autonomy as something an agent earns per workflow. Agents launch with approvals on high-risk tools and monitoring on everything, and gates are relaxed only when eval scores and production traces show the agent handling that class of action reliably. Each change in autonomy is agreed with the client and recorded with its evidence.

Autonomy in practice (Anthropic, February 2026)

Autonomy in practice (Anthropic, February 2026)
MeasureFigure
Median Claude Code turnAbout 45 seconds
99.9th percentile turn, October 2025 to January 2026Under 25 minutes to over 45 minutes
Sessions on full auto-approve, new usersAbout 20%
Sessions on full auto-approve, after 750 sessionsOver 40%
Turns interrupted, new and experienced usersAbout 5% and about 9%
Public API tool calls that appeared irreversible0.8%

Common questions

What is deployment overhang?

Deployment overhang is the gap between how much autonomy AI models could handle and how much they are given in practice, as described in Anthropic's February 2026 research on agent autonomy.

Where does the term come from?

It is used in Anthropic's post Measuring AI agent autonomy in practice (February 2026), which compares capability estimates with Claude Code and API usage data.

Does deployment overhang mean agents should get more autonomy?

Not automatically. It shows the limit is often oversight and product design rather than capability. Whether a particular agent gets more autonomy should depend on its own evals and production record.

How should a company decide how much autonomy to give an agent?

Start with approvals on high-risk actions and monitoring on everything, then relax individual gates as evidence builds up. Anthropic recommends investing in post-deployment monitoring and easy intervention rather than approval of every step.

Running agents that have to hold up?

Describe the agent, the systems it touches and what it must never get wrong. The first call is with a lab engineer, who will say plainly whether it is ready for production, and whether Motiontic is the right team.