Swamp synthesis
Type: agentic-system-analyses/types/agentic-system-synthesis.md
Bounded synthesis
From a task to executable automation
Swamp is an agent-facing automation layer, not an enclosing agent inference runtime. It supplies task briefs and skills to external harnesses, then executes the typed behaviors and workflows those agents operate. Its advertised work is reviewable, shareable and accurate operational automation. Reviewability has concrete implementation support; accuracy remains bounded by particular checks rather than demonstrated as a general outcome. The evidence is SRC-1, https://github.com/swamp-club/swamp, at commit 380e3c307164501c7d705f7707a864f4612b2785, with cutoff 2026-10-15. The whole-system functional boundary includes shipped callers, consumers and repository-local improvement workflows. Evidence consists of inspected code and source-native instructions; no target execution or causal experiment was supplied. External harness inference, conversation management, providers and hosted-service implementations are outside the frozen source. See the boundary and source register and RT-CLM-agent-workflows.
A launch loads task state, checks launchability and brief ownership, writes the brief, enrolls a harness and opens a terminal. The handoff asks the agent to read one named task and an onboarding skill. The skill directs discovery, creation, validation, target confirmation, execution and output inspection. These are agent instructions, not an enforced conversational state machine. On resume, some harnesses continue the most recent conversation in the folder; Swamp does not establish a task-specific conversation binding or own its compaction policy. RT-RTE-agent-launch, RT-BAP-agent-instructions.
The agent chooses a behavior method and arguments. Swamp resolves the definition, validates method arguments and invokes the implementation with data, vault and execution services. In source-native terminology these are “model” methods, but they need not invoke language models. A symbolic YAML workflow instead gives the scheduler next-step ownership through dependencies, conditions and expressions. Run state is saved before the started event and at material transitions; stored progress supports later continuation. A last-evaluated replay path uses an expanded retained form rather than necessarily re-evaluating current authored source. RT-RTE-behavior-run, RT-RTE-workflow-run, RT-OBJ-workflow-run.
Typed execution is not host containment. Local methods run in the host process, shell methods execute authored commands, and extension import occurs before export validation. Collective allowlisting and registration checks admit capabilities structurally, not as proven-safe code. Remote placement moves method execution to a worker, skips checks requiring worker-local state and rejects rollback-on-failure. A configured approval gate blocks ordinary waiting-run resumption, but parallel siblings may already have acted. Server authorization differs from local caller-supplied approval identity. Thus schema validation, selected preflight checks and optional gates provide local protocols, not universal human consent or rollback of external effects. RT-RTE-extension-load, RT-RTE-worker-dispatch, RT-RTE-gate-resume, EPI-RTE-preflight-check.
From results to later work
Swamp retains versioned resource/file content, workflow checkpoints, issue resources, reports and selected command events. Local content files and YAML access metadata are distinct; arbitrary payload bytes are not classified by a readable report pointing to them. Resources can contain natural-language guidance and symbolic fields with different consumers. Method argument mismatches reject execution, whereas resource-schema failure warns rather than rejects storage. MEM-OBJ-data-content, MEM-RTE-resource-storage.
The inspected read-back paths are requests: resource names and versions, CEL predicates/projections, report tags, time windows and explicit continuation. Data query is unlimited by default; count limits are not model-token budgets. A later consumer sees the selected projection, not everything physically retained. Reports deliberately omit argument values, and issue-plan review omits several rationale-bearing fields even though full JSON remains separately queryable. Retention is also conditional: infinite duration does not override finite version history, and cleanup can remove older evidence. These mechanisms support inspectable state reuse, not demonstrated recall sufficiency or automatic context injection. MEM-RTE-data-query, MEM-OBJ-method-report, MEM-OBJ-plan-main, MEM-RTE-data-cleanup.
The repository-local issue procedure gives criticism a consequential path. Caller-supplied feedback and revised plan fields are persisted, findings retain descriptions and resolution notes, and later sessions are instructed to query phase and resume the appropriate procedure. A selected local approval preflight checks that plan and review exist, their versions match and no unresolved critical/high finding remains. Those retained fields can veto approval; they do not establish the truth of criticism, actual remediation or conversational consent. The iteration method copies supplied revisions rather than reasoning about incorporation. Its latest-per-name feedback enumeration also fails to reconstruct every historical round. MEM-RTE-plan-iteration, MEM-RTE-lifecycle-checks, EPI-RTE-issue-clearance, RT-RTE-issue-admission.
This is not the only lifecycle producer. Report verification found the shipped retroactive fast_forward route, which writes a plan, empty adversarial findings, conformance entries and implementing phase for already-completed work. Consequently review existence cannot establish a criticism episode, and not every implementation transition can be assumed to traverse ordinary approval. This limits inventory completeness without invalidating the selected preflight predicate. Report-verification limit, MEM-OBJ-review-findings.
Representative delivery: commit review
The joined verify-reviews workflow is a concrete multi-step use, not a hypothetical composition of capabilities. Given a commit, setup creates a detached run-keyed worktree, initializes Swamp and restores initialization edits. Change detection obtains a merge-base diff and filename list. Code review runs unconditionally after successful detection; path guards select additional reviewers, with concurrency four. Selected reviewers receive the full diff by pathname, criteria and reachable repository files—not proof that those files were read. Each review launches a separate external Claude process with selected read tools. Only reviewer prose and shell/gate results return through the execution outputs, keeping full review conversations outside the caller's frame. Swamp does not thereby control the provider's context window or pin exact weights. RT-RTE-review-verdict, RT-CMP-agent-models, workflow account.
Pipeline failure or a non-pass parsed result rejects a review step. Recognized provider errors take precedence; otherwise the first matching line-initial pass/fail marker decides, even after a preface or before a contradictory later marker. The parser therefore accepts a protocol, not a reference answer about code quality. Cleanup is best effort. The selected workflow delivers review prose, diagnostics and run outcomes, not a merged change; it has no automatic repair-and-re-review loop. After rejection, the caller owns repair and rerun. RT-RTE-review-verdict, workflow account.
Other verification paths have different licenses. Skill evaluations test trigger/guide/tool-choice labels and external review scores; combined-rate acceptance does not imply every phase passes or generated answers are true. Verification summaries reshape recorded outcomes. The aggregate attestation derives a gate from counted step statuses and source count, permitting skips. CI checks subject, gate and supplied hashes, while stale time and missing hashes can be warnings rather than vetoes. These are evidence-relative derivation and operational admission, not semantic acceptance of every code claim or independently authenticated execution. EPI-RTE-skill-evals, EPI-RTE-report-summary, EPI-RTE-attestation-build, EPI-RTE-attestation-ci.
Learning and assessment boundary
The strongest supported learning-related contribution is automatic execution-feedback-to-durable-report production with later report/data consumers. Method and verification summaries provide a wired trace-learning route in the comparison's write-route sense, using tool traces as input. This requires neither new knowledge nor parameter updates. It does not demonstrate improved capacity. Raw audit capture, unchanged checkpoints and nonpersisted activity displays do not automatically establish additional qualifying updates. MEM-RTE-report-retrieval, EPI-RTE-report-summary, profile verification.
Retained criticism also has an instructed revision consumer, separate from the revised plan's downstream guidance. Whether criticism of a consumed theory improves capacity for future action remains unestablished: no concrete theory/criticism/revision episode or comparison closes that claim. Theory-builder condition 1 (localized theory content) is not established merely by a plan schema; condition 2 needs content-dependent decisions, not delivery; condition 3 needs a specific challenged claim, blame and changed reliance; condition 4 needs the criticism's result shaping another round. The records do not establish those four conditions for a formulated theory, although recording, revision and clearance capacities are supported. Reflection and autonomous theory building likewise remain unestablished, not evidenced absent. No exercised self-improvement is established at the shipped-system boundary: retained reports, editable capabilities and review gates do not show evidence determining an update to Swamp's own organization followed by later operation causally depending on it. MEM-RTE-plan-iteration, epistemic bounded conclusion, RT-RTE-extension-load.
For repeatable automation, the distinguishing support is symbolic scheduling, typed arguments, provenance and inspectable continuation. For issue work across sessions, it is explicit retained state plus requested reload, qualified by lossy displays and incomplete feedback accumulation. For pre-PR review, it is parallel delegated inspection with a computational rejection protocol, not measured defect detection. Execution traces would establish activation; content-linked criticism and revision episodes could establish theory use; controlled comparisons would be needed for accuracy or capacity benefit. A traced update to Swamp's own live behavior-determining artifacts and subsequent dependent operation could change the self-improvement assessment. These conclusions remain scenario-relative rather than a product ranking.
Limitations
| Limitation | Affected SRC IDs and record citations | Inspected boundary | Conclusion prevented | Evidence that would resolve it |
|---|---|---|---|---|
| No supplied execution or causal comparison | SRC-1; RT-RTE-workflow-run, RT-RTE-review-verdict, MEM-RTE-plan-iteration | Frozen implementation and instructions, not target operation | Actual success, compliance, recall quality, reviewer accuracy or beneficial learning | Inspectable runs linking inputs, retained products and later consumers; appropriate intervention comparisons for causal benefit |
| Retroactive lifecycle producer omitted from detailed inventory; report-verification limit | SRC-1; MEM-RTE-plan-iteration, MEM-OBJ-plan-main, MEM-OBJ-review-findings, EPI-OBJ-adversarial-review, RT-RTE-issue-admission; verification | Ordinary issue routes; verifier additionally inspected fast_forward in the shipped extension and skill |
Complete producer comparisons; review existence proving criticism; every implementation transition traversing ordinary approval | Accepted route-level account of retroactive production and its later consumers, plus episode evidence for actual criticism |
| Task-brief authority omitted as a separate profile unit; profile-verification limit | SRC-1; RT-OBJ-task-brief, RT-RTE-agent-launch, RT-BAP-agent-instructions; verification | Brief storage/form and launch handoff supported; partial authority inventory | Complete per-object authority and launch-versus-issue-guidance comparison | Separate classification of the retained brief's launch consumer and force |
| External runtime and service responsibilities | SRC-1; RT-CMP-agent-models, RT-RTE-agent-launch, RT-RTE-worker-dispatch | Swamp adapters, not provider weights, host compaction, hosted persistence or worker OS isolation | Model fixity, task-specific conversation restoration, universal permissions or remote durability | Frozen external implementation/contracts and deployment evidence covering those responsibilities |
| Selected checks are not universal authorization or containment | SRC-1; EPI-RTE-preflight-check, RT-RTE-gate-resume, RT-RTE-issue-admission, RT-RTE-extension-load | Local selected checks, distinct server/local paths, remote skips, host execution and import | Universal human consent, semantic safety or external-effect rollback | Route-complete authorization and deployment controls; authenticated consent and effect-specific recovery evidence |
| Retention does not guarantee complete delivery or historical incorporation | SRC-1; MEM-RTE-plan-iteration, MEM-OBJ-method-report, MEM-RTE-data-query, MEM-RTE-data-cleanup | Latest-per-name feedback enumeration, selected displays/projections and finite histories | All-feedback incorporation, complete rationale delivery or context fit | History-complete enumeration and consumer-level retrieval traces with explicit content/budget guarantees |
| Included alternatives remain incompletely inspected | SRC-1; MEM-RTE-resource-storage, MEM-RTE-report-retrieval, MEM-RTE-data-cleanup, RT-OBJ-workflow-run, EPI-RTE-skill-evals | Selected backends, custom reports, cache invalidation, authorization, hook/serve consumers, acquisition and adoption paths | Complete storage/form/curation/authority inventory, pull-only behavior, complete erasure or a system-wide absence of adaptation | Accepted records tracing the named alternatives and actual later consumers |
| Verification licenses are narrower than correctness or authenticity | SRC-1; RT-RTE-review-verdict, EPI-RTE-skill-evals, EPI-RTE-attestation-build, EPI-RTE-attestation-ci | Verdict syntax, fixture labels/scores, supplied run arithmetic and CI consistency | General semantic accuracy, valid evaluator labels, non-forgeable execution evidence or complete merge gating | Validated references, observed review outcomes, authenticated acquisition chain and repository admission-policy evidence |
No unresolved cross-report conflict was supplied. No frozen-source files were read for this synthesis; it uses the supplied members without adding evidence.