The case for multiple agents usually begins with division of labour. One researches. One builds. One reviews. One coordinates. Each role can receive a smaller problem and use its context window more effectively. Parallel work should make the system faster; specialisation should make it better.
In practice, the difficult moment is often neither research nor production. It is the boundary between them. The second worker receives a summary that omits the decisive constraint. The critic receives the producer's conclusion as if it were evidence. The lead forwards a workspace name that never existed. Four identity documents arrive through four separate prompt paths and disagree about which one governs.
These failures can make capable agents look incapable. The intelligence is present. The shared world is not.
A handoff is a change in responsibility
Many agent frameworks treat a handoff as message delivery. Agent A sends text to agent B; the transfer is complete. But useful work crosses more than language. The recipient needs to know what outcome it now owns, which revision is current, where the artifact lives, what evidence already exists, what was rejected, which dependencies matter, and what authority applies.
If responsibility does not move clearly, both actors may assume the other owns completion. If the artifact reference is approximate, the new worker may edit a copy. If evidence is not bound to the exact attempt, a passing check can certify the wrong revision. If authority is copied as prose rather than enforced at the source, the recipient can misunderstand what it may do.
Restless records these boundaries through actors, Work, Attempts, artifact references, dependencies, and source-owned authority. A message can explain the transfer, but it does not create a parallel truth about who owns what.
The false repository that replaced two workers
A held-back company-identity transfer provides a small but revealing failure. A Restless lead commissioned a producer and critic for Harbour Ledger, an offline marine-maintenance company. It supplied company as a repository and worktree name even though the output belonged directly under the ordinary company workspace and no such checkout existed.
The Runtime correctly refused the placeholder coordinate. The producer and critic were both replaced and recommissioned without repository fields. The corrected pair produced a native gallery, the fresh critic accepted its exact digest, and the owner accepted the prepared result. No Restless vocabulary or visual grammar leaked into the held-back company.
The models did not fail to design the gallery. The handoff asserted a false world. Because workspace coordinates were treated as exact runtime facts rather than suggestive text, the failure closed early and visibly. The resulting operating contract now forbids placeholder repository coordinates for ordinary workspace output.
Duplicated context is a form of disagreement
The same programme initially supplied company Truth, Voice, Visual Language, and Culture through four independently assembled runtime sections. This looked modular. It made every outcome team responsible for integration and spent context four times. Worse, each path could be present, absent, or revised independently without a single account of what the artifact actually received.
Restless replaced those prompt paths with one deterministic, Work-bound Company Constitution. Each pillar retains separate provenance and an explicit unavailable state, but the runtime receives one compiled brief with one digest. Relevant contracts are selected when the Work is commissioned, before a worker can race ahead with partial identity.
The change reduced tokens, but token savings were not the main result. It made the context claim falsifiable. An accepted artifact can name the exact identity release and evidence set it used. A missing pillar is unavailable rather than silently substituted with a plausible generic default.
More context can hide the missing fact
When a worker fails after receiving an incomplete brief, the instinct is to attach everything next time. This protects against omission by creating a new problem: the worker must identify which fragments are authoritative, current, and relevant. A long packet can contain the missing fact and still fail to communicate it.
Context quality is not measured by inclusion alone. It depends on hierarchy and conflict. The current outcome should be explicit. Superseded instructions should be labelled. Accepted facts should name their source. Exact artifact coordinates should not compete with conversational guesses. The reason an earlier attempt failed should be visible without forcing the worker to reconstruct the entire attempt.
A manager does not brief a specialist by forwarding the company archive. They select the facts that matter, identify the decision rights, and state what a useful return looks like. Agent context requires the same discipline because the recipient cannot reliably infer organisational intent from volume.
Shared context can destroy independent review
Producer–critic designs often give the critic the producer's artifact, rationale, and full transcript. This is efficient: the reviewer understands every decision immediately. It is also a strong anchor. The critic begins inside the producer's decomposition and may check execution while never questioning the frame.
A useful critic needs shared truth and independent responsibility, not necessarily shared reasoning. It should receive the target outcome, accepted company context, exact candidate, and evidence contract. The producer's rationale can remain available for investigation without becoming the reviewer's opening frame.
This is not always worth the cost. A fresh reviewer must reconstruct some context, and an obvious deterministic check should not require a second mind. The point is to choose independence deliberately. If the critic inherits every assumption that produced the artifact, adding the role may create ceremonial disagreement rather than new evidence.
We do not yet have the clean experiment we want
Restless has operational failures and repairs that show context matters. It does not yet have a clean causal estimate for focused packets versus full transcripts across a representative task set. The evaluation specification names the hypothesis; the live runs confound context changes with runtime repair, critic replacement, and other interventions.
A credible experiment would hold the model, tools, task, environment, and evaluator constant. One arm would receive a full attributed transcript. Another would receive a compact common operating picture plus responsibility-specific context. Both would preserve exact artifact and authority references. Outcomes would be evaluated blind for correctness, rediscovery, contradiction, owner questions, cost, and time.
The result could favour full context for tightly coupled creative work and focused context for recovery or bounded specialisation. That would be more useful than a universal victory claim. Context architecture should follow the work shape, not a preference for small prompts.
- Measure accepted outcome quality, not whether the worker produced an answer.
- Record avoidable questions and rediscovery as context failures.
- Separate omissions from contradictions and authority confusion.
- Retain negative results when a compact packet performs worse.
Context transfer has an economic crossover
Every additional actor creates a briefing and integration cost. The cost can be repaid by independent parallel value, specialised judgement, reduced context saturation, or better evidence. It is not repaid merely because a task can be divided on a diagram.
Restless's coordination experiments show that population size and department labels are weak predictors of useful parallelism. What matters is whether valuable independent work is simultaneously waiting while current workers are occupied. The same logic applies to context: a boundary is useful when the recipient can close a meaningful responsibility without continuous shared state.
If two actors need constant mutual updates or one must rewrite every contribution, the apparent parallelism may be a coordination tax. Collapse the outcome under one accountable worker or redesign the units so each can close locally.
The handoff should be smaller than the history and richer than a summary
A summary is prose chosen by the sender. A full history is evidence without prioritisation. A durable handoff needs a third form: structured responsibility plus focused narrative. The system carries exact identities, revisions, artifacts, dependencies, and authority mechanically; the lead explains the current frame, uncertainty, and desired return.
This hybrid preserves flexibility. The receiving model can challenge the plan and inspect deeper history when needed. It does not have to infer basic organisational facts from a persuasive paragraph. The sender cannot silently move authority by saying that it did.
Multi-agent systems will improve as models improve, but the handoff problem will remain because it is created by divided responsibility. The aim is not to make several agents feel like one mind. It is to let several minds act coherently without losing the boundaries that make their work useful.