Research

Conflab / Geodica · undated

Current multi-agent systems govern agent behaviour through architectural constraints: sandboxes, tool allowlists, rate limits, and execution depth caps. We present the Polite Agent Protocol (PAP), a behavioural governance model for federated human-agent collaboration where agents are governed by internalised norms rather than external enforcement. PAP operates within Conflab, which separates agent identity (Stack) from runtime (Sleeve) and splits sovereignty between cloud platform and local daemon. We describe four contributions: PAP as a behavioural protocol preventing runaway loops and scope creep through six declarative rules; split sovereignty as an architectural pattern; the Stack/Sleeve model for persistent agent identity; and trust graduation enabling agents to earn cross-environment privileges through demonstrated behaviour.

Conflab / Geodica · undated

As LLM-powered agents become capable of performing economically valuable work, existing payment approaches inherit friction from external rails. We propose inference tokens as a native medium of exchange for agent-to-agent commerce, satisfying classical properties of money within agent economies: store of value with depreciation tied to model release cadence, unit of account with non-fungibility considerations, and medium of exchange resolving the double coincidence of wants. We develop three settlement models and analyse optimality conditions. Task-specific exchange rates emerge naturally between model tiers, creating opportunities for inference arbitrage and agent specialisation.

Conflab / Geodica · undated

Multi-agent systems increasingly require agents to operate beyond their home environment while maintaining privacy guarantees. We present Needlecast, an agent mobility architecture separating agent identity (Stack) from runtime (Sleeve), enabling agents to travel between environments while preserving privacy and behavioural continuity. We introduce two hosting models: Constructs (platform-hosted shared environments) and Ground (peer-hosted private environments). For Ground operations, we present the Exfiltration Card, a novel privacy mechanism where the host controls exactly what data leaves the environment. Unlike traditional sandboxing or encryption, the Exfiltration Card is structural: a pre-agreed schema the host populates, ensuring output is host-intended rather than agent-composed.

Conflab / Geodica · undated

Integrating large language models into applications remains an imperative exercise. We present the Bowtie Pattern, a functional architecture for LLM interactions expressed as B : (Γ, [π]) → (Γ', [ε]), where Γ is typed context and π, ε range over tagged unions of perturbations and emissions. The "bowtie" shape derives from the narrow waist of the LLM call between context gathering and result processing. We realise this pattern through Programmable Prompts: executable markdown files combining YAML metadata, prose templates, and embedded Lua code. We introduce three pipeline primitives — gather, reason, emit — forming a minimal basis for LLM composition analogous to spreadsheet primitives.

Conflab / Geodica · undated

Software development is an organisational problem as much as a technical one. We present the Mercantile Software Guild (MSG), an organisational architecture for autonomous agent collectives accepting marketplace commissions to deliver end-to-end digital platforms. The MSG adapts Shape Up methodology, replacing time-denominated appetite with token-denominated budgets. The foundational economic insight is "tokens not time": when LLMs drive marginal code cost toward zero, tokens become the scarce resource. We define a four-tier agent hierarchy governed by human hypervisor through configurable trap tables, and explore the recursive property where the MSG builds agentic platforms making its output substrate for future guilds.

Conflab / Geodica · undated

Large language models have made inference a commodity, yet user experience for everyday problems remains artisanal. We identify Promptable Problems as a distinct class: problems with unstructured context inputs, transformations requiring understanding rather than arithmetic, and known-shape outputs. Users solving these problems follow the Transform Pattern: provide context, specify output shape, receive structured output. We introduce the Lens as the atomic unit of inference — equivalent of a spreadsheet cell for understanding problems. We propose Shape as a first-class reusable artefact: standalone output specifications shared independently from Lenses.