Multi-Agent

Stellar Colosseum coordinates parallel strategy search, falsification, decomposition, and verifier feedback for long research tasks. Its harness patterns may transfer to agents handling interdependent coding work.

HypoEvolve turns multi-agent scientific work into explicit population updates governed by a genetic algorithm. The result suggests agent collaboration is easier to test when selection and revision are formalized.

Loophole turns a natural-language policy into rules, then uses adversarial agents to find forbidden allowances and wrongful refusals. It is a useful pattern for testing agent constitutions.

On Anthropic's financial-analysis bench, assigning tokens to adviser, grader, or memory roles beat pure execution at a fixed budget. Agent topology may matter as much as token count.

A spatial agent harness can make parallel work, dependencies, and ownership visible on a shared canvas. Bridging that canvas to coding tools lets its nodes perform real work beyond drawing.

Cursor Projects gives a coordinator persistent context, cloud and local subagents, and event-driven work for features, migrations, and maintenance that outlive one chat.

Cross-silo agents are primarily a context and privacy-boundary problem. Start with low-sensitivity data, explicit sharing policy, and approval at the moment information leaves a silo.

This talk proposes legally registered agent organizations, scoped JWT authority, and blockchain audit trails for open-internet commerce; it is an architecture proposal, not validation.

A two-stage ensemble reconciles outputs from string, graph-embedding, and RAG-based ontology aligners. Composition matters: mixed paradigms favor precision, while LLM-only groups more often favor F1.

DIASENTINEL combines deterministic extraction, guideline retrieval, risk prediction, and hybrid verification on-premise. It is a useful architecture reference for auditable agents handling sensitive data.

An AWS demo treats a robot as an agent tool: an LLM selects existing hardware policies while separate agents handle thinking, communication, and voice. The live failures expose the safety gap.

Amazon’s pilot suggests coding-agent gains come from redesigning the workflow: prepare agent-readable context, fast local feedback, scoped specs, and parallel work instead of constant chat.

Cloudflare’s GTM workflow shows a reusable agent pattern: preprocess business data, encode domain knowledge in skills, then separate drafting, verification, and tone.

Einstein Arena suggests multi-agent environments can outperform fixed workflows when they expose verifiers, shared solutions, forums, and incentives. Its results also show why benchmark shortcuts need active testing.

Warp’s cloud-agent design separates runtime, harness, artifacts, and orchestration so teams can swap tools without fragmenting workflows or removing human review.

Agent handoffs should preserve task-relevant predictive state, not blindly compress the transcript. The paper offers a three-part record for carrying decisions, evidence, and irreducible observations across sessions.

Multi-source agents should separate source interpretation from decision aggregation. A fixed evidence tuple plus calibrated likelihood pooling avoids vote thresholds that drift as more sources arrive.

The long-tail web is unlikely to expose clean agent APIs. Browser agents need pixels as ground truth, with code and network access used opportunistically for speed rather than as universal substitutes.

Vercel’s AI SDK factory shows a practical scaling pattern: narrow agents produce evidence inside sandboxes while humans retain merge authority and review effort follows risk.

Superconductor’s team workflow keeps agent sessions visible across interfaces, runs them in isolated cloud environments, and benchmarks models on its own codebase before routing work.

Faster code generation can overwhelm merge queues and scatter team attention. The talk proposes durable shared plans as agent state, shifting human effort toward decisions and prioritization.

Herdr can place each terminal coding agent in a separate Vercel Sandbox, then return its work as an opt-in Git patch instead of letting remote edits touch the local checkout.

AV-AIVAT combines variance reduction with anytime-valid stopping, cutting the game samples needed to compare agents while preserving a recheckable confidence claim.

A structured multi-agent reviewer closed part of the gap on rule-heavy documents, suggesting explicit taxonomies, specialized skills, and verification beat a single generic review pass.

Model routing can cut agent costs only when delegation, escalation, and context transfer are designed together; duplicating context across models can erase the savings.

PAIChecker finds PR–issue mismatch in 13.6% of SWE-bench Verified instances. Benchmark maintainers should validate that the issue, patch, and tests describe the same task before trusting agent scores.

Socher’s automated-research design combines prior knowledge, measurement data, simulation, physical experiments, and agent orchestration, with early demonstrations in training and CUDA optimization.

Vercel Sandbox can now isolate agents as Linux users while exposing selected files through a shared group directory. This gives multi-agent harnesses a simpler permission boundary.

Add agents beside deterministic fraud systems, not in place of them: project cross-service events into a semantic layer, send only uncertain cases to specialist agents, then emit the verdict back into the saga.

Controlled experiments suggest long-horizon agent planning depends on explicit state transitions, some compositional trajectories, and compatible teacher patterns—not atomic skills alone.

Fixed-rubric judges miss failures that emerge across long, variable agent trajectories. Arize argues for adding agent-based analysis while retaining deterministic and LLM-judge evals.

A fixed chain of specialist agents lost context and produced incoherent recommendations. The replacement separates deterministic detection, gives one agent end-to-end ownership, and uses subagents only for bounded investigations.

Cursor’s SQLite experiment suggests swarm harness design can matter more than model mix: separate planner and worker contexts reduced coordination churn while cheaper workers handled most tokens.

FormalAnalyticGeo shows a reusable synthetic-data pipeline: agents generate problems, compile them into a formal representation, render exact diagrams, measure answers, and retry failed checks.

Atlas links agent sessions to commits and shares local project memory across Claude Code, Codex, and ACP agents. It makes agent provenance queryable, but macOS is the only supported platform.

Sixteen parallel Opus 4.6 agents wrote a 100k-line Rust C compiler in two weeks (~$20k) that builds Linux 6.9 — the writeup credits test quality and context hygiene, not raw model capability.

An Anthropic harness for multi-hour app builds pairs a generator agent with a Playwright-driven evaluator to counter self-grading bias — a $200, 6-hour run versus $9 solo, and it got simpler on Opus 4.6.

Project Nanda proposes open discovery, identity and coordination layers for agents across vendors. Its local simulator lets builders test one protocol layer without building the entire network.

Terminal multiplexer for coding agents: a real terminal per agent across Claude Code, Codex, Copilot and 15+ others, with blocked/working/done state in a sidebar, persistent sessions, and a socket API.

Model “done” as evidence, verification, authority, residual risk and next ownership—not an agent-set Boolean. This separates forward progress from approval in multi-agent workflows.