We Ran 38 AI Agents Through 120 Rounds of Geopolitical Simulation to Predict Lebanon's Future. Here's What They Agreed On.

The Experiment Setup

I'm Lebanese. Born in Beirut, raised through the tail end of one civil war, watched from various distances as the country went through several more crises. I run HAQQ, a legal AI company, and earlier this year we published a simulation of 72 AI agents predicting the future of legal AI. That experiment used a multi-agent social simulation engine called MiroFish. It worked surprisingly well — the agents converged on predictions that, months later, are already proving accurate.

So I thought: what if I pointed this thing at something I actually care about? Not an industry trend, but the country I come from. Lebanon has been in freefall since 2019 — currency collapse, port explosion, political paralysis, a war with Israel, and a ceasefire that may or may not hold. Everyone has opinions about what happens next. The diaspora is split between fatalism and delusional optimism. The politicians are performing business as usual. The foreign ministries are writing cables that say nothing.

I wanted a different kind of analysis. Not polling data (Lebanon hasn't had a census since 1932). Not pundit takes. I wanted to build the actual stakeholders as AI agents, give them the real information, and see what they converge on when left to argue with each other.

The Stack

The process: I wrote three seed documents. The first was a detailed profile of 20 personas — real people representing every major faction in Lebanese politics, regional power brokers, diaspora voices, historians, economists, and military actors. The second was a geopolitical landscape brief covering the situation as of early 2026: the post-ceasefire environment, the Aoun presidency, the Salam government formation, the economic crater. The third was adversarial — a set of 12 structured questions designed to force the agents into contested territory. No softballs.

MiroFish ingested these three documents into Zep Cloud's knowledge graph, which produced 90 nodes and 90 edges mapping the relationships between entities. From this graph, the engine spawned 38 distinct agents — some directly mapping to the 20 personas, others emerging as composite voices from the knowledge graph (a Sunni business owner synthesized from multiple economic nodes, for example, or a Shia academic emerging from the intersection of Hezbollah and university-related nodes).

The agents then ran for 120 rounds across simulated Twitter and Reddit platforms. Over those rounds, they produced 285 total actions: 48 original posts, 89 comments, 51 quote-posts, and 41 likes. Some posts generated heated threads. Others were ignored. The dynamics were organic — agents chose what to engage with based on their persona, their memory of previous interactions, and the evolving conversation.

Total runtime: 33 minutes. Cost: under $4 in API calls.

Let me be upfront about the limitations before we get to findings. This was a single run. Our legal AI simulation ran three parallel experiments for cross-validation. We didn't do that here, partly because this was a personal project and partly because I wanted to see if even a single run would produce signal worth reporting. It did. But take the confidence levels for what they are: the consensus of one simulation, not a statistically rigorous prediction.

The 20 Agent Personas

Each persona was built from a 400-600 word profile that included biographical background, institutional affiliation, known policy positions, rhetorical style, and — critically — their incentive structure. A simulation is only as good as its character sheets. I spent more time on these profiles than on any other part of the experiment.

The personas break into five groups:

This was the simulation's most depressing consensus — and also its most robust one. Across 120 rounds, not a single agent proposed a viable transition away from confessionalism that other agents accepted. The activists tried. Georges Sassine's agent posted early and forcefully:

"The confessional system is a cancer on Lebanon. It rewards warlords, punishes competence, and makes every citizen a hostage to their birth certificate. We didn't take to the streets in 2019 for reform within the system. We went out to burn it down."

— Georges Sassine agent (Minteshreen)

The post got engagement. Other agents liked it, quoted it, responded with solidarity. But when the conversation turned from diagnosis to prescription, it fell apart. Sassine's agent couldn't propose a transition mechanism that didn't require the cooperation of the very people who benefit from the system.

The historians were characteristically blunt. Traboulsi's agent traced confessionalism to the Mutasarrifate of 1861, arguing it was never an organic Lebanese institution but a colonial power-sharing formula that the local elites adopted because it served them. Khashan's agent agreed with the diagnosis but offered a colder prognosis: communal identity in Lebanon is real, not manufactured, and any system that doesn't account for it will fail.

The political agents were the most revealing. Geagea's agent didn't defend confessionalism ideologically — he defended it tactically. Abolishing the confessional system in a country where Hezbollah has an armed militia and the Shia demographic is growing would, in his framing, simply mean transferring power from a negotiated balance to demographic dominance. Fadlallah's agent, meanwhile, argued for reform within the system rather than abolition — but his proposed reforms happened to advantage his faction.

Joumblatt's agent captured the paradox perfectly: "Everyone wants to abolish confessionalism after they win. Nobody wants to abolish it before." The Druze position is illuminating because the Druze are a small enough community that any majoritarian system would marginalize them. Their defense of power-sharing isn't principled — it's existential.

The international agents (Macron, Pedersen) called for reform but offered no mechanism to impose it. France tried in 2020. It failed. The simulation replicated this failure with precision.

Confidence: 75%. The confessional system survives to 2035 not because it works, but because destroying it requires a level of inter-communal trust that doesn't exist and shows no sign of developing.

This finding emerged organically — it wasn't one of the seed questions. The agents independently converged on the observation that Lebanon is already functionally partitioned along sectarian lines, and the central government in Beirut is increasingly a shell that administers nothing meaningful outside the capital.

The map the agents drew (metaphorically — MiroFish doesn't do graphics) looked like this: Hezbollah exercises effective sovereignty over southern Lebanon and the Bekaa Valley. The Lebanese Forces and other Christian parties consolidate control over Mount Lebanon and parts of the North. Sunni-majority areas, left leaderless after the Hariri withdrawal, are gravitating toward Gulf patronage for basic services. Beirut is contested ground that functions as a shared administrative space but is governed by no single faction.

The LAF agents pushed back hard on this framing:

"The Lebanese Armed Forces opposes any form of partition, whether formal or informal. The army deploys across all Lebanese territory and answers to the central government. We are the last national institution, and we will not accept a narrative that treats the state as already dissolved."

— Colonel Samer Nassar agent (LAF)

But the historians offered a different read. Traboulsi's agent noted that Lebanon's borders were drawn by the French in 1920 to create a Christian-majority state, and that the country has never had a stable internal consensus on what it is. The current de facto partition is just the latest iteration of a pattern that predates independence. "You can't partition what was never truly unified," he argued.

Khashan's agent went further, noting that even during the "golden age" of the 1960s, state authority in the South and Bekaa was minimal. The only thing that changed after 2006 is that Hezbollah's parallel state became more visible and more capable.

The economic agents made a crucial point: soft partition is happening despite economic integration, not because of it. Lebanon's economy runs through Beirut's port, airport, and banking system (what's left of it). The regions cannot function as independent economic units. This means the partition dynamic creates a worst-of-both-worlds scenario: political fragmentation with economic interdependence, ensuring that no region can cleanly separate but no central authority can effectively govern.

Maha Yahya's agent framed this as "the governance vacuum" — a situation where formal institutions exist on paper, foreign embassies still present credentials to a president in Baabda Palace, but actual service delivery, security provision, and dispute resolution happen through confessional networks, not state mechanisms.

The simulation's verdict: soft partition deepens through 2035 unless an unexpected shock (major external intervention or a genuine national unity crisis) reverses the trend. The probability of formal federalism being enacted is low (15%) because no faction wants to give up its claim to the whole, but the reality on the ground will increasingly resemble a federal arrangement without the legal framework.

This was the topic that broke the simulation's usual tone of measured analysis. The agents got heated. The thread on external intervention became the longest and most contentious in the entire 120-round run, with agents from every faction weighing in and several engaging in multi-round arguments.

The core observation is not new: every major crisis in Lebanese history has been resolved (or prolonged) by external actors. The 1958 crisis ended with American marines. The 1975-1990 civil war ended with the Saudi-brokered Taif Accord. The 2005 Cedar Revolution was triggered by a Syrian assassination and enabled by American and French pressure. The 2006 war was between Hezbollah and Israel, with Iran providing weapons and the UN brokering the ceasefire. The 2020 port explosion didn't even produce a crisis resolution — just French posturing that went nowhere. The 2024 war ended with an American-mediated ceasefire.

The Lebanese agents across the political spectrum were resigned to this pattern. Sassine's agent called it "the original sin of Lebanese statehood." Joumblatt's agent, ever practical, noted that a small country between Israel, Syria, and the Mediterranean has never had the option of true sovereignty and never will.

But the external agents were the revealing ones. The Ron Dermer agent was, by far, the most provocative actor in the entire simulation:

"A weak and divided Lebanon serves our strategic interests best. A unified, sovereign Lebanon with a competent military would either become an adversary or fall under the influence of actors hostile to Israel. The current situation — a country too fragmented to threaten us but stable enough not to collapse into chaos on our border — is, from a purely strategic standpoint, close to optimal."

— Ron Dermer agent (Israel)

This comment received 7 dislikes — the most negative reaction of any single interaction in the simulation. Every Lebanese agent, regardless of faction, reacted negatively. Even Fadlallah's Hezbollah agent and Geagea's Lebanese Forces agent — who agree on almost nothing — both condemned it. For a brief moment, the simulation showed what might actually unite the Lebanese: a common enemy saying the quiet part out loud.

The Saudi agent (Prince Faisal) took a softer line but revealed a similar logic: Saudi engagement in Lebanon is calibrated to counter Iran, not to build Lebanese institutions. If the Iranian threat receded, Saudi interest in Lebanon would drop to near zero. The French agent was the most sympathetic but also the most impotent — grand rhetoric about Lebanon's "special relationship" with France paired with a consistent inability to deliver meaningful outcomes.

The Iranian agent was the most strategically coherent. Bagheri Kani's agent framed Iran's relationship with Hezbollah not as interference but as partnership, and argued that Iran's influence in Lebanon is a consequence of Lebanese state failure, not its cause. "If the Lebanese state had protected the Shia south from Israeli aggression, there would have been no need for the resistance," he posted. This generated a furious response from the Christian and Sunni agents, but the historians quietly noted he wasn't entirely wrong about the causal chain.

Geir Pedersen's UN agent offered the normative framework: sovereignty, UNSCR 1701, international law. But when pressed by other agents on enforcement mechanisms, the UN agent had nothing. 1701 has been violated for nearly two decades by both sides. The UN's role in Lebanon has devolved from peacekeeping to witness-bearing.

The simulation's conclusion: external actors will remain the decisive factor in Lebanon's trajectory. No purely internal dynamic has the power to overcome the external interference structure. This is not a prediction — it's a diagnosis of a permanent condition.

Father Fadi Tabet's agent raised the Petit Liban concept early in the simulation. This is the idea, rooted in 19th-century Maronite political thought, that Lebanon's "natural" borders are smaller than the current state — essentially Mount Lebanon and parts of the coast, a Christian-majority entity that existed as the Mutasarrifate from 1861 to 1918.

The concept has emotional power for many Lebanese Christians. The Greater Lebanon created by the French in 1920 added Muslim-majority areas (Tripoli, the Bekaa, the South, Beirut's Sunni quarters) to the Christian core, and the demographic math has been moving against the Christians ever since. The last census (1932) showed a thin Christian majority. Nobody believes that's still the case. Estimates now put Christians at around 30-33% of the population, and the emigration numbers are devastating: the simulation's historical agents estimated that Christians could drop below 25% by 2035.

So the appeal of Petit Liban is understandable: if you can't win the demographic race, redraw the map. Father Tabet's agent framed it in cultural terms — preserving the Maronite heritage, the cedars, the monasteries, a way of life under threat.

The historians contextualized it. Traboulsi pointed out that the Petit Liban concept was rejected by the very Maronite leaders who built modern Lebanon, because they wanted Beirut's port, the Bekaa's agriculture, and Tripoli's commerce. A smaller Lebanon would be economically unviable. Khashan added that the demographics within Mount Lebanon itself are shifting — Druze and Shia populations have been growing even in what would be the Petit Liban territory.

The LAF agent rejected it on institutional grounds. Colonel Nassar's agent was firm: the military's oath is to Lebanon's current borders, and any partition proposal undermines the army's raison d'etre.

But it was the economists who delivered the killing blow:

"Lebanon's economy is integrated. The port of Beirut handles goods for the Bekaa and the North. The banking system — what remains of it — is headquartered in Beirut but services the entire country. The agricultural supply chains run from the Bekaa to the coast. Partition would impoverish all regions. You cannot draw borders through a city of 2 million people and expect anything other than economic collapse on both sides."

— Sami Halabi agent (economist)

Dr. Nadia Fawaz's agent ran the numbers. A Petit Liban of roughly 4,000 km2 with perhaps 1.5 million people would have no port (Beirut straddles the dividing line), no agriculture (the Bekaa is outside), limited water resources, and a GDP roughly equivalent to a mid-sized European city — without the institutional capacity. It would be entirely dependent on the same diaspora remittances that currently prop up the whole country, but now flowing to a fraction of the territory.

The simulation's consensus: Petit Liban will continue to surface as a rhetorical device, particularly when Christian emigration numbers are in the news or when Hezbollah asserts itself politically. But no major political party will formally propose it (25% probability by 2030), and even if proposed, it has zero chance of implementation. It's therapy, not policy.

The economic threads in the simulation were the most technical and, in some ways, the most despairing. Every agent who engaged with economic questions — whether economist, politician, or activist — converged on the same structural diagnosis: the banking sector's collapse is the root cause of the economic crisis, and no recovery is possible without a resolution of the banking sector's losses.

The numbers are staggering. Lebanon's GDP fell from approximately $55 billion in 2018 to under $20 billion by 2021. The banking sector has an estimated $72 billion hole — the gap between what banks owe depositors and what they actually have. The Lebanese pound, officially pegged at 1,507.5 to the dollar for decades, collapsed to over 89,000 before a new rate was set. Inflation was north of 200% at its peak. The poverty rate went from 25% to over 80%.

The IMF deal that was supposed to provide a framework for recovery has been stalled since 2022. The Sami Halabi agent was precise about why: the IMF requires banking sector restructuring that would force politically connected depositors and bank shareholders to take losses. No political faction will impose this because the major depositors are the political class themselves and their networks.

"The banking crisis is not an economic problem. It's a political problem with economic symptoms. The banks are owned by the political elite or their proxies. The depositors who matter — the ones whose calls get returned — have already transferred their money out. What's left in the banks is the savings of ordinary Lebanese who are being slowly and deliberately wiped out through inflation."

— Sami Halabi agent (economist)

The Fawaz agent added global context: no country that has experienced a banking collapse of this magnitude has recovered to pre-crisis GDP levels in less than 15 years without either massive external aid (think Marshall Plan scale) or a commodity boom (think Gulf states after oil price recoveries). Lebanon has neither. The only comparable cases in the Middle East are Iraq and Syria, and both had oil to leverage.

The one lifeline that every agent acknowledged: diaspora remittances. At $7-8 billion per year, remittances now constitute roughly 30-35% of GDP. This flow has kept the country from complete societal collapse. But remittances are consumption money, not investment capital. They keep people fed but don't build factories or fund infrastructure.

The political agents' response to the economic crisis was revealing. Fadlallah's Hezbollah agent talked about "resistance economy" — a vaguely defined concept of economic self-sufficiency that economists dismissed as meaningless. Geagea's agent blamed corruption (true but incomplete). Joumblatt's agent didn't even try to propose solutions — his economic interventions were about protecting Druze communities from the worst effects, not fixing the national problem.

Nawaf Salam's agent was the most reform-oriented, pushing for IMF compliance and banking restructuring. But the simulation showed what happens to reformers in Lebanon: they get blocked by parliament, undermined by the central bank, and abandoned by international backers when the political cost gets too high. The Macron agent's support for Salam was enthusiastic in principle and absent in practice — France won't sanction its own banking interests that are intertwined with Lebanese banks.

Simulation projection: GDP per capita returns to 2018 levels by 2035 with only 15% probability. The most likely path is a prolonged, low-grade recovery to perhaps 60-70% of pre-crisis levels, sustained by remittances and a small service economy. For the average Lebanese earning in local currency, this means a decade of diminished living standards with no clear end point.

This was the topic the agents were most divided on — and the division itself was informative.

The historical pattern is hard to argue with. Lebanon had a civil war from 1975-1990. A major Israeli invasion in 2006. A sectarian crisis in 2008. The 2019 economic collapse (which, while not armed, involved the same structural fissures). And the 2024 Israeli military operations. That's roughly one major crisis every 10-15 years, and the intervals are arguably getting shorter.

The historians presented the structural argument: every power-sharing formula in Lebanon has an expiration date. The National Pact (1943) lasted until 1958. Its renewal lasted until 1975. The Taif Accord (1990) lasted until 2005, and then entered a slow deterioration that culminated in 2024. Each formula "solves" the previous crisis by creating a new equilibrium that contains the seeds of the next one.

Khashan's agent put it bluntly: "The question is not whether there will be another conflict. The question is what the trigger will be, and which external actor will be involved."

The Hezbollah agents were the most interesting on this question. Fadlallah's agent showed pragmatic tendencies that might surprise people who read only Western media coverage. His agent argued that Hezbollah's political wing genuinely wants integration into the Lebanese state — not as a surrender of arms, but as an expansion of the party's legitimate political power. The 2024 war was devastating for Hezbollah's military infrastructure, and the political wing sees an opportunity to shift the party's center of gravity from military to political.

But there was a catch. When pressed by other agents on whether Hezbollah would disarm as a condition for full political integration, Fadlallah's agent was clear: not without Iran's explicit blessing, and Iran has no incentive to bless it. The Bagheri Kani agent confirmed this without stating it directly, framing the resistance's arms as "the community's insurance policy" rather than as an Iranian strategic asset.

The LAF agents presented the security establishment's view: the current ceasefire is fragile, the army doesn't have the capacity to replace Hezbollah in the south, and any premature attempt to enforce a state monopoly on violence could trigger exactly the conflict it's supposed to prevent.

The Israeli agent was clinical: Israel will continue to conduct operations against what it perceives as Hezbollah military reconstitution, regardless of Lebanese sovereignty concerns. This creates a standing provocation that could escalate at any time.

The simulation's agents split on probability estimates for another major armed conflict within 10 years, with assessments ranging from 40% to 65%. The median was around 55%. This is high enough to be alarming but not high enough to be deterministic. The agents who argued for the lower end (40%) pointed to Hezbollah's post-2024 weakness and the general exhaustion of the Lebanese population. Those arguing for the higher end (65%) pointed to the structural drivers: external interference, arms accumulation, unresolved territorial disputes, and the historical cycle.

One scenario that multiple agents flagged: a conflict triggered not by Hezbollah-Israel dynamics but by internal Lebanese tensions. As the confessional balance shifts and the central government weakens, the risk of inter-communal violence increases. The 2008 events — when Hezbollah briefly seized parts of Beirut — were a preview. The simulation suggested that if federalism discussions become serious, the process itself could become a conflict trigger, as each faction tries to establish "facts on the ground" before borders are drawn.

The Prediction Scorecard

We seeded the simulation with 12 structured questions drawn from the adversarial brief. These were designed to be answerable with a probability estimate and to cover the full range of Lebanon's political, economic, security, and social trajectory. Here's where the agents landed:

The scorecard tells a story even without the commentary. Look at the "Low" column: census, disarmament, IMF reforms, federalism, GDP recovery, international stability classification. All the things that would represent genuine progress are in the single digits or low teens. The only "High" confidence predictions are the persistence of a broken system and the continued decline of the Christian population. The medium-confidence items are all about conflict and disruption.

This is a simulation telling us: the status quo persists, the problems deepen, and the risks of violent disruption are significant. If you wanted optimism, the agents didn't have any to offer.

What the Simulation Missed

I want to be honest about the limitations, because overstating the reliability of a single-run simulation would be irresponsible.

No parallel runs. The legal AI experiment ran three parallel simulations and cross-validated findings. This experiment ran once. That means we can't distinguish between robust findings (those that would emerge across multiple runs) and artifacts (those that appeared due to specific random seeds or agent interaction patterns). The high-confidence findings likely hold up — they're also the ones most consistent with conventional expert analysis. The medium-confidence findings are where parallel runs would matter most.

Language issues. MiroFish's default prompts include Chinese-language instructions, and the report agent generated some sections in Mandarin Chinese before we caught it. This is a known issue with the engine's localization and didn't affect the simulation dynamics — agents interacted in English — but the automated report output required manual translation and cleanup. If you're using MiroFish for non-Chinese experiments, be prepared to deal with this.

Diplomatic agent bias. Most agents were too measured. Real political actors are more extreme, more petty, more opportunistic than even well-designed AI personas. The simulation's agents argued like policy analysts more than politicians. The notable exception was the Ron Dermer agent, which was genuinely provocative and generated the most authentic-feeling reactions from other agents. More adversarial agents would improve output quality — the engine needs personas who will say the uncomfortable things that real actors think but don't say publicly.

Missing actors. Twenty personas is a lot, but Lebanon's political ecosystem includes dozens of relevant actors we didn't model. There's no Syrian government agent, despite Syria's ongoing influence. No Turkish agent, despite Ankara's growing interest in Sunni Lebanese networks. No Palestinian faction agents, despite 400,000+ Palestinian refugees living in Lebanon. No representative of the Alawite community. No business oligarch persona beyond the Salameh context. Each missing actor is a dynamic that the simulation didn't capture.

Static seed data. The simulation's knowledge was frozen at the time of the seed documents (early 2026). Real events are moving fast — the ceasefire situation, government formation, banking discussions. A simulation that ingested real-time news would produce different (and potentially more accurate) results.

No emotional modeling. The agents argue rationally based on their incentives. Real Lebanese politics is deeply emotional — blood feuds, personal vendettas, sectarian trauma passed across generations. The simulation captures the structural logic but misses the affective dimension. This matters because in Lebanon, emotional dynamics often override rational calculation. A leader who "should" compromise based on incentive analysis might refuse because his father was killed by the other faction's militia in 1983.

The Bottom Line: Is Lebanon Safe to Live In?

This is the question everyone asks. The diaspora asks it before booking flights home. The foreigners ask it before accepting NGO postings. The Lebanese who stayed ask it to justify their decision to themselves.

The simulation's answer is more nuanced than any of them want to hear.

Physical safety: Post-ceasefire, Lebanon is physically safer than it was in 2024. The south is quiet, albeit devastated. Beirut is functioning. The risk of random violence is low by regional standards (far lower than Syria, comparable to Jordan). But "safe" is a relative term in a country where the next war is a coin flip within a decade.

Economic survival: Devastating for anyone earning in Lebanese pounds. The minimum wage covers a few days of groceries. Dollarization is nearly complete in the private sector, creating a two-tier society. For people earning in USD (diaspora returnees, remote workers, NGO staff), Lebanon is paradoxically livable — cheap real estate, good food, functioning social fabric, a level of cultural richness that the numbers don't capture. For the average Lebanese on a local salary, it's grinding survival with no visible end date.

Institutional hollowness: The state provides almost nothing. Electricity is still on generators and private solar. Water is trucked. Public schools and hospitals are barely functional. The university system is collapsing. The courts don't work. Property rights are uncertain. If you need the state for anything — a birth certificate, a building permit, a legal judgment — expect corruption, delays, and arbitrariness.

The agents' consensus scenario — "muddling through" at 35-40% probability — is both the most likely and the most insidious outcome. It means Lebanon won't die. It won't collapse into a Somalia-style failed state. The restaurants will stay open in Gemmayzeh. The nightclubs will play music in Jounieh. The ski resorts will operate in the Cedars. From the outside, from the Instagram feed of a diaspora Lebanese visiting in summer, it will look fine. Even charming.

But underneath the surface, the country will continue to hollow out. The best and brightest will continue to leave. The institutions will continue to decay. The informal economy will continue to replace the formal one. The sectarian leaders will continue to divide the shrinking pie. And every 10-15 years, the accumulated tensions will find a release valve, and the cycle will repeat.

Lebanon won't die. But it won't heal either.

That's the simulation's verdict, and it's mine too. I wish the agents had surprised me. They didn't. They told me what I already knew but didn't want to accept: that the country I come from is not on a trajectory toward recovery. It's on a trajectory toward managed decline, interrupted periodically by violence, sustained by the financial lifeline of a diaspora that left because the country couldn't support them.

If you're Lebanese reading this, I'm not telling you to give up hope. I'm telling you to direct your hope toward achievable things: your family, your community, your immediate circle. The macro picture is grim. The micro picture — the neighbor who helps without being asked, the generator owner who runs a tab for the family who can't pay, the teacher who shows up to a school that hasn't paid her in months — that's where Lebanon actually lives. The simulation doesn't capture that. No model does.

Running This Yourself

MiroFish is open source and available at github.com/666ghj/MiroFish. It's a multi-agent social simulation engine built for exactly this kind of experiment: give it seed documents, let it build a knowledge graph, spawn agents, and watch them interact across simulated social media platforms.

The Lebanon simulation required:

• 3 seed documents (~5,000 words total): persona profiles, geopolitical brief, adversarial questions
• An OpenRouter API key (for Gemini 2.0 Flash — cost under $4 for the full run)
• A Zep Cloud account (for graph-based persistent memory)
• About 33 minutes of patience

You can view the interactive simulation output, including the full agent interaction history and knowledge graph visualization, at lebanon.dashable.dev.

If you run a parallel experiment — especially one with different persona compositions or more adversarial agents — I'd love to see the results. The whole point of open-source simulation is that one run is a hypothesis. Multiple runs are data. Lebanon needs better data.

Experiment by: Stephane Boghossian — Founder, HAQQ Legal AI
Engine: MiroFish (open source)
Interactive results: lebanon.dashable.dev
Related: 72 AI Agents Predicted the Future of Legal AI