The agentic internet

Headline findings

• 2029

  6464%

  of requests come from AI task agents, while only 25% of internet users regularly use one.

• 2030

  8181%

  of requests come from agents acting for people: roughly four in five.

• 2036

  9696%

  of requests come from AI task agents, and 99% of all requests are automated.

• 2026 → 2036

  ×2626

  total requests, while direct human activity is still about half its 2026 level (47%).

• AI task agents pass half of all public-facing requests between 2028 and 2029 in the planning case (34.7% in 2028, 63.8% in 2029), while regular adopters are 12% and 25% of internet users.
• Direct human activity keeps falling in absolute terms, not only in share: down 13% by 2029 and 47% of its 2026 level by 2036, inside a total that grows about 26.1-fold.
• Human requests fall below 1% of the total by 2036 (0.85%); all automation, agents plus other bots and crawlers, reaches 99.15%.
• The result leans hardest on the assumed 2026 starting share and on how efficient agent traffic becomes; both are shown under sensitivity.

Annual results, planning case

The planning case is mildly bullish: slightly above our own central judgment, not a statistically established percentile. Both it and the central case assume agents become reliable mass-market tools.

Assumptions by year

Scenarios

Each scenario changes how fast people adopt agents and how much work each adopter delegates, and holds everything else. They are alternative assumptions, not confidence intervals, and carry no assigned probabilities. Our central judgment is our own subjective reference, not an industry consensus.

Hyper-adoption reaches 99% AI task-agent traffic in 2033. The central and planning paths do not reach 99% task agents by 2036.

Sensitivity

The assumed 2026 task-agent share is an especially weak anchor: no global, deduplicated series yet measures regular autonomous-agent users, external workload per user or the task-agent request share. Much more efficient APIs could also cut request shares sharply with unchanged adoption and delegation, which is why 99% agent traffic is not a necessary condition for an agent-first economy.

Where the 2036 workload comes from

Net workload per adopter (K) multiplies four judgments, each relative to 2026. It counts expected executed work, not every possible child in an unrestricted tree of agents, and business requests are additional calls, never a second count of one exchange.

Method

Scope

We count public-facing application requests to websites and internet-facing APIs once per request. Network packets, bulk and video byte measurements, model reasoning and internal sub-agent messages that never cross that boundary are excluded. AI task-agent traffic is a mechanism category, whatever the intent behind it, good or bad; other automation covers conventional bots and AI crawlers. The model is not a measure of economic value, purchases, jobs, human attention or bandwidth.

Starting point

Thales reports 47% human and 53% automated traffic for its 2025 sample (S01). We use that split only as normalization context, and we assume the automated 53 points divide into 1 point of AI task-agent traffic and 52 points of other automation. We also assume regular agent users start at 0.8% of internet users. Neither figure is measured.

Equations

With t years after 2026, p the share of internet users who regularly use agents, J jobs per adopter, F mean executed worker fan-out, E external requests per worker (all relative to 2026) and b independent business-agent requests per consumer-agent request:

K = J × F × E × (1 + b) / (1 + b₂₀₂₆)
N = 1.02^t
A = A₂₀₂₆ × N × (p / p₂₀₂₆) × K
H = H₂₀₂₆ × N × (1 − 0.75p) / (1 − 0.75p₂₀₂₆)
O = O₂₀₂₆ × N × 1.035^t
Total = A + H + O; each share = category / Total; all automated = (A + O) / Total

Human-origin activity falls by up to 75% at full adoption, so not all human activity disappears. Internet use grows 2% a year, which changes volumes but not shares, and other automation grows 3.5% a year faster than users. No unlimited replication or all-to-all communication between agents is assumed.

The 2036 endpoint

In 2036, p = 0.82, J = 6, F = 4, E = 0.5 and b = 1, so K = 6 × 4 × 0.5 × (2 / 1.2) = 20. That gives 95.72% AI task agents, 3.43% other automation and 0.85% human requests: 99.15% automated, in a total of 2,610.5 against 100 in 2026. Direct human activity is 47% of its own 2026 level, despite being under 1% of the larger total.

Checks

We checked that categories sum to 100%, that parameters stay within bounds, recomputed the 2036 endpoint independently with decimal arithmetic, compared the workbook's formulas with an independent recomputation, and scanned the workbook for formula errors. The model is designed for monthly revision; inspect the sensitivity rather than reading decimal places as confidence.

Evidence ledger