SwarmWorld synthesis
Type: agentic-system-analyses/types/agentic-system-synthesis.md
Bounded synthesis
From configured world to local research decisions
SwarmWorld is an enclosing runtime for societies of model agents conducting research in a shared simulated world. The shipped Python drivers, under the historical name biofoundry, own scheduling, state, action admission, world effects and recording. Its declared purpose is to study discovery, testing, exchange, inheritance and embodiment of technologies—not to provide a general coding agent or certify real materials. This account uses implementation and doctrine/design evidence from SRC-1, https://github.com/lamm-mit/SwarmWorld, at commit 6af7ae9fa36d98b07b0492cf139658e8af1f6eab, with cutoff 2026-10-15. The whole-system source boundary includes headless, live, human-play, environment, replay and research-tooling paths; specialist coverage remains selective. No target execution, provider call, episode trace or paper dataset was supplied or produced (boundary, Source register, EPI-CLM-research-environment).
The representative workflow is a headless LLM episode because one shipped driver connects configuration to terminal delivery. It creates the simulation and recorder, then alternates model planning with simulator ticks until the configured horizon. A software scheduler fans out agent-specific calls; there is no lead model delegating to subordinate models and synthesizing their answers. Each selected agent receives local observations, an action manual, selected retained evidence and its latest research state. Concurrent calls propose private state and bounded action sequences against a world that has not yet stepped. The policy binds returns to agent IDs and replaces queues; the simulator later consumes one action per agent per tick (workflow account, RT-RTE-model-planning, RT-OBJ-action-queues).
Context is rebuilt as a fresh message pair, not an accumulating provider conversation. Latest research state is delivered separately from ranked private retrieval. Private queries combine prior state, pending requests and local observations; public, spatial, empirical and program views have different selectors. Optional attention controls remove retrieved entries before shrinking observation lists. These omissions do not delete stored originals, but excerpts can lose necessary conditions, diagnostics can count selected records that never reach the model, and an irreducible causal core can leave the character budget exceeded. Delivery is wired; reliance on a particular record is not demonstrated (MEM-OBJ-research-state, MEM-RTE-context-selection, EPI-RTE-memory-selection).
From proposed plans to measured effects
There are two admission boundaries. First, local JSON/schema checking admits the representation of a model plan, filters research-state citations for availability and writes the accepted state before queue commitment. This does not check whether citations support the hypothesis. The local maximum also need not enforce a smaller provider-configured plan limit. Second, simulator dispatch checks treatment permissions and action-specific resources, targets and constraints. A schema-valid action can therefore be rejected by the world. Rejection feedback can inform later planning, but neither a completed tick nor the instruction against invented observations establishes truthful prose or successful scientific repair (RT-RTE-model-planning, RT-RTE-world-stepping, RT-CLM-grounded-science).
Measurements and eligibility checks have distinct authority. Passive sensing and INSPECT retain locality- and time-bounded simulator evidence. Observed or possessed resources can license recipe inputs. TEST reads a fabricated batch's simulated properties and utility; its private result omits the internal target-threshold flag. When experience reuse is enabled, the source recipe remains paired with the measured outcome. Proposal-time evaluation is another check: combined designs can receive a comparison against leave-one-inherited-material-out recipes. That contrast concerns material inclusion under the simulator's utility function, not the independent contribution of a publication, communication or model intelligence. These implemented checks produce evidence or grant operational eligibility; they do not accept a model's explanatory theory as true (EPI-RTE-sensor-evidence, EPI-RTE-batch-measurement, EPI-RTE-recipe-admission, EPI-RTE-proposal-evaluation, MEM-OBJ-research-mission).
Experience also crosses social boundaries. Teaching can copy retained evidence or program knowledge to nearby agents; publication adds archive and spatial access surfaces. Personally grounded resource names in publication text can grant later combined-design eligibility. This is stronger than merely storing prose, but narrower than endorsing its assertions. Pending-request and fulfillment edges route work and result receipts without checking whether a successful action satisfied the request's natural-language meaning. Private notebooks, the public archive, spatial deposits and the request ledger consequently cannot be treated as one uniformly warranted knowledge base (MEM-RTE-social-retention, EPI-RTE-insight-circulation, MEM-OBJ-message-records, MEM-BAP-evidence-admission).
Agents can change persistent artifact behavior through validated symbolic programs. BUILD, WRITE_PROGRAM and FORK_PROGRAM converge on DSL validation and installation; fork-specific ancestry checks do not govern every installation. The straight-line VM has bounded vocabulary and instructions, not arbitrary Python or shell access. Installed instructions determine actuator requests on later artifact ticks, whereas the same instructions in a model's skill view supply descriptive context. Validation establishes executable constraints, not usefulness, host isolation or a single aggregate actuator budget across repeated instructions. Registry history, installed code and per-agent knowledge indices remain distinct. A skill's verified flag means measured exposure and can be transferred; it does not certify competence. Configured inheritance preserves selected measured skills across replacement while private and empirical memories reset (RT-RTE-program-install, RT-BAP-program-execution, MEM-OBJ-program-library, MEM-OBJ-program-skills, MEM-RTE-cultural-inheritance).
Recovery, delivery and later evaluation
Invalid model output or provider errors yield WAIT proposals. Optional outage freezing has a narrower, driver-owned contract: any qualifying transient failure in a scheduled batch restores scheduled memories, preserves old queues and returns noncommitment; the shipped episode driver waits and retries without stepping. This is not limited to total outage, does not cover every error class, and gives no eventual-completion guarantee. The convenience policy actions() route does not honor noncommitment in the same way. Simulator action rejection separately provides no general whole-tick rollback (RT-RTE-model-planning, RT-RTE-world-stepping, workflow account).
Normal delivery means final recording, recorder closure and return of a metrics summary and trace path—not scientific success. Optional frozen-society assays deep-copy a discovery state, remove active agents and measure simulated persistence under configured conditions, including artifact-knockout contrasts where enabled. Their inspected episode consumer records results rather than using them to approve live plans or choose a successor controller. Offline research tooling also includes endpoint metric selection and embedding/cache mechanisms; the live finding must not become a system-wide absence of automatic selection (RT-RTE-ecology-assay, workflow account, report verification).
Replay is a separate consumer. It restores committed research-state records and steps recorded actions, subject to compatible revision, scenario, tick and digest checks. It does not ask the model to decide again. Selector statistics are outside the inspected authoritative memory serialization, and retrieval scoring/use evolution is not restored. Even successful digest agreement would therefore establish specified fixed-action state agreement, not complete adaptive continuation or scientific correctness (RT-RTE-recorded-replay, memory replay annotation, MEM-OBJ-event-trace).
Direct human, scripted and environment controllers bypass model-output admission but still reach simulator checks. The research-oracle is a privileged fixed-recipe controller, not an answer oracle consulted by the LLM. Separate scenario authoring changes data and guidance outside episodes; loader validation and hashing do not establish enforced human review or in-world self-modification (RT-RTE-direct-control, RT-RTE-scenario-admission).
What the retained changes establish
The strongest supported adaptive contribution is a wired experience-to-later-consumer chain. Research-state continuation, empirical summaries and measured skill indices retain experience for later context or admission. With adaptive retrieval enabled, selection and accepted citations during planning, followed by action outcomes during stepping, update persistent counters that alter future ranking. These are genuine behavior-shaping memory routes without requiring a weight update. They meet the comparison profile's trace-learning criterion, but that criterion does not establish improved capacity for future action. Citation-and-success credit is associative; no supplied comparison identifies its benefit or the effect of a particular memory component (MEM-RTE-retrieval-feedback, MEM-RTE-evidence-retention, MEM-OBJ-research-state, profile verification).
The independent theory and self-change conclusions are correspondingly bounded:
- Theory-builder condition 1: afforded by localized hypothesis fields and whole-state replacement, not demonstrated by an instantiated theory. Condition 2: uninspected; delivery does not demonstrate content-dependent decisions. Condition 3: uninspected; no supplied claim/result/blame/revision instance establishes content-directed criticism. Condition 4: uninspected; persistence alone does not show criticism shaping the next round. Whole research-state fields are addressable, but the structure of actual assumptions is uninspected (RT-RTE-model-planning).
- Criticism improving future capacity: uninspected. Retained outcomes and adaptive access establish a partial mechanism, not a criticism-led improvement or its attribution (RT-RTE-model-planning, MEM-RTE-retrieval-feedback).
- Reflection and autonomy: requested goal/progress self-description and automatic scheduling do not establish both causal directions of a reflective representation or computation performing every internal theory-building role. Reflective and autonomous theory building remain uninspected, independently of software-owned orchestration (RT-RTE-model-planning).
- Self-improvement: adaptive counters change selection organization, and artifact installation changes simulated products, but exercised self-improvement of the bounded runtime is unestablished. Objective-aligned determination and subsequent live behavior causally depending on the update are not demonstrated. No success or enclosing-runtime improvement follows from installation alone. External weight mutability is separately uninspected (EPI-RTE-memory-selection, RT-RTE-program-install, RT-CMP-configured-model).
For studying how locally informed model societies propose, measure, exchange and embody candidates under controlled simulation rules, the inspected architecture supplies the relevant operational loop. It does not establish that a society completed useful discovery, learned a better explanation or transferred a mechanism to physical reality. Linked episode evidence showing delivered context, candidate content, checks, criticism, retained revisions and subsequent decisions would change the exercised-property assessment. Independent variation of retrieval feedback, communication or inheritance under a declared objective would be needed for component-benefit attribution. A complete compatible checkpoint including selector state would change the adaptive-continuation assessment. These are evidence requirements, not inferred failures of the implemented research environment (EPI-CLM-research-environment, RT-RTE-model-planning, MEM-RTE-retrieval-feedback, memory replay annotation).
Limitations
| Limitation | Affected SRC IDs and record citations | Inspected boundary | Conclusion prevented | Evidence that would resolve it |
|---|---|---|---|---|
| No supplied execution or paper data | SRC-1; RT-RTE-model-planning, RT-RTE-ecology-assay, boundary | Frozen source implementation and doctrine; no target execution | Successful episodes, recovery, replay, actual reliance, scientific benefit and causal paper results | Inspectable compatible episode artifacts and scoped intervention comparisons |
| Offline producer/consumer and evaluator inventory remains partial; carried report-verification limit | SRC-1; report limit, EPI-RTE-proposal-evaluation, RT-RTE-ecology-assay, MEM-OBJ-event-trace | Reports reserve study, statistical attribution, dense-representation provenance, visualization and depiction consumers; verifier additionally inspected endpoint metric selection and embedding/cache paths | System-complete evaluator, representation or automatic-selection comparisons; especially absence of offline selection inferred from the live assay path | Accepted mechanism-level records tracing offline producers, retained representations, selection criteria and downstream consumers |
| Profile units group mechanisms and omit some local classifications; carried profile-verification limit | SRC-1; profile limit, RT-OBJ-artifact-program, MEM-OBJ-program-skills | Partial comparison inventory does not separately enumerate installed program substrate or emit coarse verified-status/recency selection for program views; authority and direction units group consumers | Complete object counts, complete per-selector value sets and one-to-one route comparisons; emitted existence findings remain supported | Finer accepted units separating installed slots, program selectors and individual consumers/operations |
| Semantic warrant and theory use are not demonstrated | SRC-1; RT-RTE-model-planning, EPI-RTE-memory-selection, EPI-RTE-insight-circulation | Schema/ID admission, excerpts, publication and later delivery; no instantiated criticism chain; further offline routes unassessed | Citation support, faithful delivery of every condition, accepted explanatory novelty, exercised theory-building, or a system-wide absence of such routes | Actual input/output propositions, support checks and a linked criticism/revision/later-decision chain; broader records for system-wide conclusions |
| Adaptive-memory reconstruction is incomplete | SRC-1; RT-RTE-recorded-replay, MEM-OBJ-retrieval-stats, replay annotation | Committed research-state restoration and recorded-action stepping; selector evolution not restored or covered by inspected serialization | Complete adaptive-checkpoint equivalence or resumed model-society behavior from digest agreement | Selector-state persistence/restoration and a compatible continuation comparison |
| Simulator checks do not validate physical mechanisms | SRC-1; EPI-RTE-batch-measurement, EPI-RTE-proposal-evaluation, RT-RTE-ecology-assay | Surrogate measurements, arithmetic utility and matched simulated contrasts; full physics not independently audited | Real-material accuracy, explanatory truth or attribution finer than the actual contrast | Validated encoding/physics and external measurements; independently varied components for causal attribution |
| Provider internals, exact weights and deployment controls remain outside supplied evidence | SRC-1; RT-CMP-configured-model, RT-RTE-direct-control, RT-RTE-program-install | Adapter requests, controller routes and bounded simulated execution, not service internals or deployed security | Weight fixity, provider retention compliance, host sandboxing or deployment authentication assurance | Pinned serving identity, provider-side evidence and inspected deployment enforcement |
No additional frozen-source reads were needed for this synthesis; its findings use the supplied members and their retained evidence.