delegation.md

Delegation to experts and specialists

How a collective decides which specialist acts, in biology and in current agentic systems, and which of those mechanisms can be encoded as a pure fold over an attributed transcript with fixed-point integer arithmetic. This is the reading behind ../specs/expert-delegation.md and ADR 0007.

The gap

Before this work the library could express "who is here" and "who spoke", and could not express "who knows". RosterMember is {id, name}; the only expertise-shaped field, AgentThreshold.affinity, was host-supplied and never written by the library; cross-desk referral routed to a named target's home desk with nothing choosing the target by expertise; the responder ladder's selector saw a role string and nothing derived from the transcript.

Biology and collective behaviour

For each mechanism: the rule an individual runs on local information, then what a transcript would have to represent to fold it.

Fixed response thresholds (Bonabeau, Theraulaz & Deneubourg 1996; 1998). Individual i carries a private threshold θᵢⱼ per task j; task j emits a public stimulus sⱼ that rises while undone; engage with probability sⱼⁿ/(sⱼⁿ + θᵢⱼⁿ), n = 2. Two castes with θ₁ < θ₂ produce the whole allocation with no allocator and no messages. Fold: the attention market's argmax over urge - threshold already is this. The missing half is that stimulus is not per task: salience is per trace and relevance is a static lookup.

Threshold reinforcement (Theraulaz, Bonabeau & Deneubourg 1998). θ falls by ξ per unit time on the task and rises by φ off it, clamped to [θ_min, θ_max]. Specialists emerge from a homogeneous population and the specialisation reverses when the specialist leaves. The clamp is load-bearing; MAX–MIN Ant System (Stützle & Hoos 2000) uses the same clamp against premature convergence. Fold: affinity as a count of own deposits on a topic, clamped, which is commutative; iterated updates are not.

Foraging for work (Tofts & Franks 1992; Franks & Tofts 1994). No thresholds: wander, do work where found, drift to an adjacent zone otherwise. Fold: topic adjacency is co-citation, a graph over cites. Probably over-engineering.

Interaction-rate encoding (Gordon 1996; Gordon & Mehdiabadi 1999; Greene & Gordon 2007; Prabhakar, Dektar & Gordon 2012). A harvester ant decides whether to forage from the rate of brief contacts with returning foragers, not from what any contact says. Fold: the arrival rate of a topic's traces over a window is itself an allocation signal, and the library does not measure it.

Honeybee recruitment. The waggle dance advertises a site with intensity proportional to the scout's own absolute private estimate (Seeley & Visscher 2008); roughly 5% of bees ever visit more than one site, so nobody compares. Advertising decays by a fixed absolute amount per return (Seeley & Visscher 2003), a quality-proportional survival filter. The tremble dance (Seeley 1992) fires when a forager waits more than about 50 s to unload: it recruits a different caste and suppresses recruitment of more foragers. A load signal that reallocates labour across castes, triggered by a locally measured queueing delay. The stop signal (Seeley et al. 2012) is cross-inhibition targeted at dancers for other sites, already implemented here as !object. The tandem-run to carrying switch at quorum (Pratt et al. 2002; Pratt 2005) moves the colony from a mode that preserves independent judgement to one that does not, and maps onto Deliberate → Commit. Fold: an intensity must be earned from cites, never parsed from prose, or it is free to inflate.

Teaching and informed leaders. Tandem running is teaching by the strict criteria, at about four times the solo journey time (Franks & Richardson 2006): an expert who transfers knowledge is slower than one who acts. Couzin, Krause, Franks & Levin (2005): a small informed minority steers a group whose naive members cannot tell who is informed, and the proportion needed shrinks with group size. Fold: a directory does not need to be broadcast to work; it only needs to bias the argmax.

Transactive memory. Wegner (1986; 1995): a group's memory is the directory, not the contents, with directory updating, information allocation and retrieval coordination. Lewis (2003) validates specialisation, credibility and coordination; credibility is what lets a retriever trust an owner instead of re-deriving. Hollingshead (1998; 2000): teams rely on diffuse cues (role label) early and specific cues (observed experience) as they gain shared history; the effect of diffuse cues falls with experience. A host-supplied role string is a diffuse cue; a folded record of grounded deposits is a specific one.

Hidden profiles. Stasser & Titus (1985): groups discuss what everyone knows and fail to surface uniquely held items. Stasser & Stewart (1992): the failure largely disappears under a "demonstrably correct answer" framing. Stewart & Stasser (1995) and Stasser, Stewart & Wittenbaum (1995): public expert role assignment raises unique-item sampling, modestly (about 29% to 34%; Lu, Yuan & McLeod 2012 meta-analysis). Rule: announce my domain; when a question falls in another's announced domain, ask them rather than answer; when it falls in mine, contribute even unasked.

Switching cost. Leighton, Charbonneau & Dornhaus (2017): the interval between acts is shorter when an individual repeats the same task, in every active worker group and independently of whether the worker is a specialist. Division of labour buys the avoidance of switching delay, not better workers. Fold: a fixed penalty when the topic a member would speak on differs from its last one; SPEAK_COST's mirror image.

Quorum and pooling. Sumpter & Pratt (2009): quorum responses need k ≥ 2 and actively suppress early social influence below the threshold. Marshall et al. (2019): simple majority is frequently sub-optimal.

Agentic systems, 2023 to 2026

system	who acts next	signal	fan-out	failure mode
AutoGen 0.2 GroupChat	auto (LLM names next speaker), round-robin, manual, callable	LLM over full history	sequential	transcript read per turn
AutoGen FSM / 0.4 SelectorGroupChat	allowed-transitions graph ∩ LLM; selector_func, candidate_func	static graph + LLM	sequential	wrong edges deadlock silently
LangGraph supervisor, hierarchical teams	supervisor node emits Command(goto)	LLM, static topology	sequential	supervisor is a context bottleneck
CrewAI hierarchical	manager calls "delegate to coworker" by role string	LLM matching task to role	sequential	delegation loops
OpenAI Swarm / Agents SDK	handoff as a tool; control fully transferred	LLM tool choice	sequential	no return edge
MetaGPT	publish/subscribe on cause_by	static subscriptions	broadcast	bounded only by disjoint watch sets
CAMEL, ChatDev	fixed dyad; waterfall SOP	topology	sequential	role flipping, loops
Anthropic research system (2025)	orchestrator spawns 3–5 workers	LLM decomposition	parallel	~15× tokens; usage explains ~80% of variance
Claude Code subagents and skills	task matched against description	LLM over a short index	either	the description is the router; misroutes are silent

Learned and scored selection: DyLAN (Liu et al. 2023) computes an Agent Importance Score by backward aggregation of peer ratings and keeps the top-k, the closest published analogue to a citation-folded directory. AgentVerse recruits generated expert personas (unfalsifiable). GPTSwarm (Zhuge et al. 2024) learns who talks to whom. MasRouter (2025) jointly picks collaboration mode, roles and per-agent model, with up to 52% overhead reduction. AgentRouter (2025) routes over a knowledge graph. Bala & Shah (2026) frame routing as set-valued prediction and find supervised routers beat zero-shot LLM routing: "ask a model who should answer" is the baseline, not the ceiling. Mixture-of-Agents (Wang et al. 2024) improves even from worse auxiliary answers and is rarely reported against a matched-budget control.

Cost-aware cascades: FrugalGPT (cheapest first, escalate on low scorer confidence, a quorum on a confidence signal), RouteLLM, Hybrid LLM, RouterBench, IRT-Router. Predictive routing decides from the query; cascade routing needs an abstention signal the answerer emits. KnowNo (Ren et al. 2023) is the principled abstention trigger: ask for help when the conformal prediction set is not a singleton. Expert personas (ExpertPrompting, multi-expert prompting) are diffuse cues and can trade accuracy for alignment.

The honest half

1. The compute confound. Token usage explains most multi-agent gains; matched-budget self-consistency is the control and this repository's vote arm stays as it is.
2. MAST (Cemri et al. 2025): 14 failure modes from 1600+ traces across seven frameworks, 41% to 87% failure rates. Step repetition (15.7%), unawareness of termination (12.4%) and disobeying the task (11.8%) dominate; role violation is 1.5%. An expertise layer buys little against that distribution and adds surface for the first three.
3. Convergence. Conformity rises with a peer's apparent capability, and a directory amplifies apparent capability by construction.
4. Directory circularity. DyLAN's score, Claude Code's description match and any transcript-folded affinity share a defect: who spoke becomes who is thought to know. That is an information cascade with a routing table.
5. Every added verb here has lost. !refute cost 7 points and dropped below the vote control; require_evidential cost 26. A new verb is a new way to spend turns on meta-conversation.

Synthesis: six mechanisms, ranked

1. A folded transactive-memory directory feeding BidReason::Knows. Wegner + Lewis + DyLAN. Specialisation = decayed grounded deposits on a topic; credibility = decayed citations by others minus objections; host affinity as a prior. No new verb. Benchmark: a hidden profile with heterogeneous expertise; the acceptance criterion is already written in the shared-medium spec. Might lose because: uniform-expertise benches give it nothing to route on (expected gain there is zero); circularity; mutual citation is gameable; a 6–7 turn episode is short for any estimator.
2. Render the directory into the room. Hutchins's speed bugs; Stasser's expert-role assignment. Host pastes Directory::lines() into the prompt. Small honest effect; also renders the prior.
3. !defer as abstention and cost-aware escalation. Tremble dance; FrugalGPT. Not a wasted turn but one bit of high-value directory evidence. Might lose because a turn costs 100% of a turn where a cheap model call costs about 1% of an expensive one; miscalibrated deferral is the dominant failure of escalation systems; defer chains are a step-repetition surface.
4. Threshold reinforcement plus a switching cost. No wire change. Hard to justify at 6.75 turns per episode without cross-episode persistence the crate deliberately does not own.
5. Expertise-directed referral. Couzin; LangGraph supervisor. The directory is folded from this conversation, so another desk's members weigh zero unless a host supplies cross-desk priors, and a returned answer is an unrebuttable cascade seed.
6. !offer with an earned quality signal. Waggle dance. Pure meta; the repository's record on added verbs argues against it.

Acceptance criterion, written before the numbers

The mechanism must be able to lose and the loss must be published; the bench arm and scenario family land before the mechanism and are fixed before tuning; vote gets the same turn budget and results are reported at equal turns and as turns-to-decision at equal accuracy; a mechanism that helps hidden profiles but costs more than two points on the uniform 5000-room bench ships off by default; directory circularity is reported as the rank correlation between directory weight and speech share; the predicted result on the uniform bench is zero.

References

• Bonabeau, Theraulaz & Deneubourg, Proc. R. Soc. B 263:1565 (1996); Bull. Math. Biol. 60:753 (1998).
• Theraulaz, Bonabeau & Deneubourg, Proc. R. Soc. B 265:327 (1998).
• Robinson, Annu. Rev. Entomol. 37:637 (1992). Tofts & Franks, TREE 7:346 (1992); Franks & Tofts, Anim. Behav. 48:470 (1994). Tripet & Nonacs, Ethology 110:863 (2004).
• Gordon, Nature 380:121 (1996); Gordon & Mehdiabadi, Behav. Ecol. Sociobiol. 45:370 (1999); Greene & Gordon, Behav. Ecol. 18:451 (2007); Prabhakar, Dektar & Gordon, PLoS Comput. Biol. (2012).
• Seeley, Behav. Ecol. Sociobiol. 31:375 (1992); Seeley & Visscher, Behav. Ecol. Sociobiol. 54:511 (2003); J. Exp. Biol. 211:3691 (2008); Seeley et al., Science 335:108 (2012).
• Pratt, Mallon, Sumpter & Franks, Behav. Ecol. Sociobiol. 52:117 (2002); Pratt, Behav. Ecol. 16:488 (2005).
• Franks & Richardson, Nature 439:153 (2006). Couzin, Krause, Franks & Levin, Nature 433:513 (2005).
• Leighton, Charbonneau & Dornhaus, Behav. Ecol. 28:319 (2017).
• Sumpter & Pratt, Phil. Trans. R. Soc. B 364:743 (2009); Marshall et al., eLife 8:e40368 (2019); Kao & Couzin, Proc. R. Soc. B 281:20133305 (2014).
• Czaczkes, Grüter & Ratnieks, Annu. Rev. Entomol. 60:581 (2015). Stützle & Hoos, FGCS 16:889 (2000).
• Wegner, in Mullen & Goethals, Theories of Group Behavior (1986); Social Cognition 13:319 (1995). Lewis, J. Appl. Psychol. 88:587 (2003). Hollingshead, Group Processes & Intergroup Relations 3:257 (2000).
• Stasser & Titus, JPSP 48:1467 (1985); Stasser & Stewart, JPSP 63:426 (1992); Stewart & Stasser, JPSP 69:619 (1995); Stasser, Stewart & Wittenbaum, JESP 31:244 (1995); Lu, Yuan & McLeod, PSPR (2012).
• Hutchins, Cognition in the Wild (1995).
• AutoGen speaker selection: microsoft.github.io/autogen (0.2 docs, FSM GroupChat blog, 2024). LangGraph supervisor: github.com/langchain-ai/langgraph-supervisor-py. CrewAI hierarchical process: docs.crewai.com. OpenAI Swarm: github.com/openai/swarm. MetaGPT communication docs. CAMEL arXiv:2303.17760; ChatDev arXiv:2307.07924.
• Anthropic, "How we built our multi-agent research system" (2025). Claude Code subagents documentation.
• Wang et al., Mixture-of-Agents, arXiv:2406.04692. Liu et al., DyLAN, arXiv:2310.02170. Chen et al., AgentVerse, arXiv:2308.10848. Zhuge et al., GPTSwarm, ICML 2024. MasRouter arXiv:2502.11133. AgentRouter arXiv:2510.05445. Bala & Shah arXiv:2606.28925 (2026). Ishibashi & Nishimura, Self-Organized Agents, arXiv:2404.02183.
• Chen, Zaharia & Zou, FrugalGPT, arXiv:2305.05176. Ong et al., RouteLLM, arXiv:2406.18665. Ding et al., Hybrid LLM, arXiv:2404.14618. Hu et al., RouterBench, arXiv:2403.12031. RouterEval arXiv:2503.10657.
• Xu et al., ExpertPrompting, arXiv:2305.14688. Multi-expert prompting, arXiv:2411.00492. Ren et al., KnowNo, arXiv:2307.01928.
• Cemri et al., "Why Do Multi-Agent LLM Systems Fail?", arXiv:2503.13657. "Multi-LLM-Agents Debate: Performance, Efficiency, and Scaling Challenges", ICLR 2025 blogpost track.

Items with 2026 arXiv identifiers surfaced through search snippets; only arXiv:2606.28925 was fetched directly. Verify the rest against the primary PDF before citing them elsewhere.