We ran Britain's net zero future through 19,758 AI agents

Building the discourse engine

The architecture ran two tiers of AI agents in parallel on a Hetzner server over 100 hours of wall-clock time. Fifty-eight core institutional agents — government bodies, energy companies, NGOs, politicians, academics, media outlets — ran on Qwen-72B via Featherless.ai. The remaining 19,700 public personas, representing a cross-section of UK demographics by age, profession, region, income, and political leaning, ran on a local Qwen 2.5 3B model at zero API cost. The dual-model approach kept the simulation economically viable while enabling a population size no single-model run could match.

Each of the 300 rounds represented approximately one month of discourse. Thirteen milestone events were scheduled across the 25 years: the GB Energy Act and Heat Pump Mandate in 2025, Clean Power 2030 achieved, the petrol and diesel car ban in 2035, the tightest-ever Carbon Budget 6 in 2037, CCUS at scale in 2043, 100GW offshore wind in 2045, and Net Zero 2050 at round 295. Agents reacted to each milestone in character, within the constraints of their configured stance, influence weight, and activity level.

Sentiment stayed high. Suspiciously high.

Across all five eras, positive sentiment ran at 88–94%. The earliest period, 2025–2030, was the most optimistic at 95%, coinciding with the GB Energy Act and the first wave of clean power infrastructure. The most subdued era was 2040–2045 at 88%, when deep decarbonisation challenges — carbon capture at scale, hydrogen HGVs, residual emissions with no easy technical fix — came into focus.

The persistent positivity has a clear structural explanation: the institutional agents who drove most of the discourse were configured as predominantly supportive of net zero, reflecting real-world survey data on UK public and institutional opinion. This is a simulation of how discourse behaves given those stances, not a poll of public sentiment. What the slight downward drift from 95% to 88% between 2025 and 2045 does capture is something recognisable: the difference between policy optimism and implementation fatigue. The first five years are the easiest to be enthusiastic about. The last ten are where the hard residuals live.

The petrol ban broke the record

The highest activity in the simulation, by some margin, came in the 2035–2040 era: 482 Twitter actions and 265 Reddit actions — nearly double the volume of the 2030–2035 clean power delivery period. The trigger was the petrol and diesel car ban, confirmed at round 120. Cost-of-living concerns also peaked in this era. Regional inequality, which ran as a persistent undercurrent across all five eras, sharpened as the transition costs became lived rather than projected.

This is worth pausing on. Of all thirteen milestone events — Clean Power 2030, 100GW offshore wind, net zero achieved — it was the car ban that generated the loudest simulated response. The reason is not surprising to anyone who has watched UK climate policy in the real world: the car ban is tangible. It sits in people's driveways. Every household with a petrol car has to reckon with it. It is not a carbon budget percentage or a grid capacity figure — it is a deadline that tells you what you will and will not be allowed to buy. Policy that lands in the driveway lands differently to policy that lands in the energy mix.

Who actually drives the conversation

Of 19,758 agents, the top 20 by action count were exclusively institutional. National Grid ESO led on Twitter with 42 actions, followed by ECIU (35) and Shell UK (34). Not a single public persona appeared in the top 20. On Reddit, where the comment-to-post ratio reached 10.6:1 — 274 comments on 11 seed posts — public engagement was almost entirely reactive: responding to institutional content, not setting new agendas.

This is not a design flaw. It reflects something accurate about how climate discourse works online: institutional voices with large followings, professional communications resources, and high configured activity levels set the frame, and the public responds. What the simulation surfaces is the degree to which that framing gap is structural — it does not close even when you populate the simulation with nearly 20,000 members of the public.

The undercurrent that never went away

Regional inequality and cost-of-living concerns appeared in every single simulated era. Not as the dominant topic — Carbon and Climate and Electric Vehicles held those positions throughout — but as a consistent secondary signal across all 25 years: 70 mentions in 2025–2030, 10 in 2030–2035, 50 in 2035–2040, 11 in 2040–2045, 9 in the final era. The numbers are small relative to the headline themes, but their persistence is the point. A simulation that shows regional inequality resolving as net zero milestones are achieved would tell one kind of story. A simulation that shows it refusing to disappear — even in the celebratory final rounds — tells a harder one.

What this simulation is and is not

This is a discourse simulator, not a climate model. It does not predict emissions trajectories or technology adoption curves. What it does — and what this run demonstrated — is model how different types of voices interact, amplify, and respond to policy milestones across time. The limitations are documented: sentiment analysis was keyword-based rather than semantic, Reddit was still running at data extraction (279 of 300 rounds complete), and the local 3B model used for public personas showed some generic social media patterns in middle rounds.

But the core findings — that 25 years of synthetic UK climate discourse would be dominated by institutional voices, concentrated around the most tangible policy moments, with regional equity concerns running as an unresolved undercurrent — feel true in a way that is hard to dismiss. That is what simulations are for: not to predict, but to surface the structural features of a story before you have to live through it.