/projects/agentswarm
A Python GPT-5.5 orchestration framework that fans work out to 35 scoped sub-agents — direct, council, research and security — then compresses findings upward into one synthesized decision, cutting input tokens ~74%.
Large multi-agent systems can waste context and produce noisy consensus. The challenge is assigning work to enough agents to get coverage while keeping total token use, synthesis complexity, and decision quality under control.
Context-Optimized AgentSwarm routes tasks across direct, council, research, and security agents, allocates budgets, audits context use, compresses findings hierarchically, and synthesizes one decision.
The hard parts are choosing the right agent topology, preventing context bloat, compressing without losing dissent, and making budget usage measurable enough to improve over time.
The framework focuses on context economics as an architecture constraint. It treats token budget, agent role, compression path, and final synthesis as explicit design variables.
This demonstrates Stefan's expertise in swarm orchestration, context engineering, GPT-5.5-style routing experiments, budget-aware agents, council patterns, and security-review subteams.
Context-Optimized AgentSwarm is about context economics. It asks how many agents should work, what each one should see, how findings should be compressed, and how final synthesis should preserve dissent.
Direct, council, research, and security agents have different jobs. Explicit topology prevents every agent from becoming a generic reviewer with the same bloated prompt.
Token budget is treated as an architectural constraint. Allocating and auditing context by role makes cost, latency, and quality tradeoffs visible.
Hierarchical compression keeps child-agent findings useful without passing every raw token upward. The challenge is preserving minority concerns and evidence while reducing context.
Security and council agents create structured adversarial pressure. That makes the final answer more robust than a majority-vote consensus.
Fits research orchestration well: quick agent definitions, clear data classes, and simple mock/live runtime switching.
The Responses API gives the swarm structured outputs, model control, and consistent accounting for token-budget experiments.
Classifies the task and decides what each sub-agent is allowed to see.
Caps input and output tokens by role so the swarm is not brute-forcing context.
Removes duplicate context and sensitive material before packets are sent.
Split architecture, implementation, metrics, memory, and output-contract work into focused lanes.
Pressure-test factuality, reasoning, product fit, and quality from multiple angles.
Evaluate evidence, tradeoffs, cost, latency, and reliability in parallel.
Probe injection, leakage, permissions, sensitive operations, and hallucination exposure.
Compresses child findings upward so the final synthesis sees the signal, not every token.
Produces the answer and leaves a decision-memory record for the next run.