AgentsAutonomy & tool use
Flower Labs' Lorenzo Sani argues that agent trajectories are the next target for federated learning
Flower Labs research scientist Lorenzo Sani, whose team has pre-trained LLMs of up to 7B parameters with federated learning, argues on Practical AI that agent trajectories are the next private data to train on across organisations: prompts, tool calls and retrieved context.
From the Practical AI podcast episode of November 13, 2026 and cited sources: Federated averaging now works for LLM pre-training. [3] [4] [5]
SOURCES[3] The Future of LLM Pre-training is Federated… arxiv.org[4] Photon: Federated LLM Pre-Training (arXiv 2411.02908) arxiv.org[5] Photon, MLSys 2025 proceedings proceedings.mlsys.org
AI-assisted, not human-edited. Methodology Corrections Who owns this site
- Original source
- Practical AI podcast episode, independent perspective
- Brief by
- AT Pulse research desk AI-assisted; publisher and contact on the desk page
- When
- November 13, 2026 podcast episode published
- Covered
- Nov 14 edition
- Evidence
- 10 sources across 5 sites; transcript sourced
- 1Practical AI: Implementing Federated Learning to enhance…share.transistor.fm
- 2Transcript: Implementing Federated Learning to enhance…share.transistor.fm
- 3The Future of LLM Pre-training is Federated…arxiv.org
- 4Photon: Federated LLM Pre-Training (arXiv 2411.02908)arxiv.org
- 5Photon, MLSys 2025 proceedingsproceedings.mlsys.org
- 6Introducing Endeavor 1.1flower.ai
QuietHow it’s measured
YouTube video · 36 views
Reach blends public engagement (views, likes, comments, shares, stars) from the platforms where the story or its original source lives: YouTube, Hacker News, GitHub, Substack and similar. A video and the article written from its transcript share one score, taken from the video. Platforms with no data are left out rather than counted as zero, and the age of a story never lowers its score.

Federated averaging now works for LLM pre-training. Sani and colleagues trained models of up to 7B parameters. Their Photon system, presented at MLSys 2025, reports using 64–512× less communication than standard distributed training, but these are the authors' own figures and no one else has replicated them yet. [3] [4] [5]
Sani describes agent trajectories as tree-shaped. Sub-agents and tool calls branch out to file systems, retrieval systems or MCP servers, and their results are then compacted back into the main thread. That means private context builds up as runs get longer. He says enterprises need these full traces to find where agents fail. [1] [2]
Early research suggests federated training on agent trajectories can work but is difficult. Fed-SE reports about 10% higher average task success than FedIT across five environments. To get there, it trained only on high-return trajectories, because sparse binary rewards destabilise local updates. [9]
federated pre-training has been shown at 7B scale, and one academic paper reports gains from federated agent training.
Read the full assessment
Implication: organisations could improve agents on sensitive traces without pooling the data, but privacy protections have to be engineered, not assumed.
Executive brief
Lorenzo Sani, a Flower Labs research scientist and Cambridge PhD candidate, co-led research showing that LLMs up to 7B parameters can be pre-trained with federated learning. On Practical AI, published 13 November 2026, he made the case that agent trajectories should be the next target for federated training: the prompts, tool calls, retrieved files and branching sub-agent context that agents build up while they work. Flower Labs' newest commercial model, Endeavor 1.1, is proprietary, and its public pages say nothing about federated training. The podcast is commentary and presents no new evaluation. The underlying research claims come from the authors themselves.
Read the full section
Lorenzo Sani, a Flower Labs research scientist and Cambridge PhD candidate, co-led research showing that LLMs up to 7B parameters can be pre-trained with federated learning. On Practical AI, published 13 November 2026, he made the case that agent trajectories should be the next target for federated training: the prompts, tool calls, retrieved files and branching sub-agent context that agents build up while they work. One tension stands out. Flower Labs' newest commercial model, Endeavor 1.1, is proprietary, and its public pages say nothing about federated training. The podcast is commentary and presents no new evaluation. The underlying research claims come from the authors themselves.
What changed and event timeline
DeepMind publishes DiLoCo
DiLoCo is a federated-averaging variant for poorly connected compute clusters. The paper reports parity with synchronous training on C4 using 500× less communication. Sani names DeepMind's work as an early precursor ().
Position paper argues for federated pre-training
Sani et al. claim federated LLM training holds up under data heterogeneity, hardware heterogeneity and partial participation, with models up to 7B trained ().
Fed-SE applies federated learning to agent trajectories
This independent paper trains agents on filtered, successful trajectories across environments and aggregates their updates in a low-rank subspace ().
Flower launches Endeavor 1.1
Flower Labs releases a proprietary model and markets readable thinking traces that customers can use to train their own models ().
Podcast frames federated learning for agents
Sani argues that private trajectory data is how organisations will find agent failure modes and fix them ().
Capabilities and access
- Photon: open source and built on the Flower framework. Models up to 7B have been trained with it (Photon; Flower blog).
- Endeavor 1.1: not open-weight. Flower says it uses open-weight models plus its earlier 7B model, "Lizzy".
- The episode introduces no new model or release.
Read the full section
- Photon: open source and built on the Flower framework. Models up to 7B have been trained with it (Photon; Flower blog).
- Endeavor 1.1: not open-weight. It is available on request through Flower-managed hosting, private cloud or on-premises deployment, with EU inference options. Flower says it uses open-weight models plus its earlier 7B model, "Lizzy". Parameter count, architecture and licence are not published (Endeavor page).
- The episode introduces no new model or release.
Technical analysis for researchers and developers
- Photon's recipe: federated averaging with small client batches and very high learning rates.
- Trajectory structure: Sani describes trajectories as tree-shaped. The amount of private context grows as the run gets longer (transcript, 19:33–21:38).
- Implementation difficulty: Fed-SE finds that sparse binary rewards leave most trajectories with little learning signal, which destabilises local updates.
Read the full section
- Photon's recipe: federated averaging with small client batches and very high learning rates. The authors attribute its data efficiency to federated averaging's tolerance of hyperparameter choices (Photon).
- Trajectory structure: Sani describes trajectories as tree-shaped. Tool calls and sub-agents branch off, fetch context from file systems, retrieval (RAG) or MCP servers, and are then compacted back into the main thread. The amount of private context grows as the run gets longer (transcript, 19:33–21:38).
- Implementation difficulty: Fed-SE finds that sparse binary rewards leave most trajectories with little learning signal, which destabilises local updates. Its fixes are filtering for high-return trajectories and parameter-efficient fine-tuning (Fed-SE).
Claims and evidence
- Photon's results: the paper reports 7B-scale federated training with better perplexity than centralised training, 35% faster wall-clock time, and convergence twice as fast as DiLoCo.
- Endeavor 1.1 benchmarks: Flower reports GPQA 94.4, HumanEval 99.9 and a composite score of 97.18, which it says beats rival frontier models.
- Fed-SE: reports about 10% higher average task success than FedIT across five environments (Fed-SE).
Read the full section
- Photon's results: the paper reports 7B-scale federated training with better perplexity than centralised training, 35% faster wall-clock time, and convergence twice as fast as DiLoCo. These are the authors' own numbers and have not been replicated independently (Photon).
- Endeavor 1.1 benchmarks: Flower reports GPQA 94.4, HumanEval 99.9 and a composite score of 97.18, which it says beats rival frontier models. These are vendor figures with no published evaluation protocol (Flower blog).
- Fed-SE: reports about 10% higher average task success than FedIT across five environments (Fed-SE).
- Corroboration: none of the reviewed sources independently corroborates the episode's argument that federated learning for agents works.
Context and prior work
Federated learning started as a way to train on data held on phones and in hospitals without collecting it centrally. DiLoCo and Photon repurposed federated averaging for low-communication LLM pre-training (DiLoCo). Sani's own route into the field was through medical-privacy research in Bologna, and he was an early Flower user (transcript, 02:03).
Read the full section
Federated learning started as a way to train on data held on phones and in hospitals without collecting it centrally. DiLoCo and Photon repurposed federated averaging for low-communication LLM pre-training (DiLoCo). Sani's own route into the field was through medical-privacy research in Bologna, and he was an early Flower user (transcript, 02:03). The shift from pre-training data to agent trajectories follows the industry's move toward agent systems.
Limitations, safety and contested findings
- Privacy is not guaranteed by default. Sani acknowledges that privacy guarantees in federated learning are still an open research problem.
- Shared updates can leak data. Earlier work rebuilt training inputs from shared gradients with token-level accuracy for text (Deep Leakage from Gradients, NeurIPS 2019).
- Adoption is uneven. Sani says some companies abandon federated learning after finding it impractical for their use case (transcript, 10:09).
Read the full section
- Privacy is not guaranteed by default. Sani acknowledges that privacy guarantees in federated learning are still an open research problem. Keeping data on-site is the only fixed protection; differential privacy works in some settings and not others (transcript, 08:51–10:09).
- Shared updates can leak data. Earlier work rebuilt training inputs from shared gradients with token-level accuracy for text (Deep Leakage from Gradients, NeurIPS 2019).
- Adoption is uneven. Sani says some companies abandon federated learning after finding it impractical for their use case (transcript, 10:09).
- Prompt injection came up informally. At his Halloween party, guests tried to game self-hosted judging agents with signs asking for high scores. He says it did not work, but this was a party game, not a security test (transcript, 24:43).
Business and practitioner implications
- Treat agent traces as an asset. Sani says enterprises need full trajectories to diagnose failures, and closed providers often don't disclose them.
- Cross-silo collaboration is the near-term use case. Hospitals and other data holders can pool learning across sites with current pipelines (transcript, 37:49–39:47).
- Check the privacy budget before trusting the label. Before treating "federated" as privacy compliance, ask whether differential privacy or secure aggregation is actually applied.
Read the full section
- Treat agent traces as an asset. Sani says enterprises need full trajectories to diagnose failures, and closed providers often don't disclose them. He favours open-weight models as the starting point for privacy-heavy applications (transcript, 12:01; 33:03–34:49).
- Cross-silo collaboration is the near-term use case. Hospitals and other data holders can pool learning across sites with current pipelines (transcript, 37:49–39:47).
- Check the privacy budget before trusting the label. Before treating "federated" as privacy compliance, ask whether differential privacy or secure aggregation is actually applied.
- Expect to own outputs. Owning model outputs and being allowed to train on them is becoming a vendor selling point (Endeavor page).
The source trail.
Sources (10)
Practical AI: Implementing Federated Learning to enhance Agentic AI
share.transistor.fmTranscript: Implementing Federated Learning to enhance Agentic AI
published_transcript
share.transistor.fm