Prime Agent synthesis
Type: agentic-system-analyses/types/agentic-system-synthesis.md
Bounded synthesis
From request to local effects
Prime Agent is a shipped coding and research runtime, not only a memory API. This account covers its Rust application, Python kernel, continual harness and bundled skills at commit 33da15a2c15c4e903c372f69ae8e1156d77d53da, with an applicability cutoff of 2026-10-14. SRC-1 supplies inspected implementation and attributed design instructions. No target code, tests, model services or examples were executed; no observed-run or causal-experiment evidence is available. Whole-system scope includes the material shipped callers and retained-state consumers, but inspection is selective rather than an exhaustive audit. External providers, user workloads, live search results and linked research papers are not frozen evidence. See the boundary and Source register.
The ordinary headless route admits a caller's prompt, constructs or restores a session, assembles model context and requests an assistant response. The model proposes text and Python calls. Rust owns tool dispatch, result admission and the decision to request another turn. Python executes in a persistent namespace and can reach shell commands, imported skills, MCP clients and host callbacks. A normal answer is terminal only when steering, follow-up, child work and goal/autonomous continuation do not require more work. Print mode then emits text or JSON events and cleans up inline descendants. These are implemented control paths, not evidence of successful task completion. RT-RTE-print-prompt, RT-RTE-python-effects.
Interactive and ACP clients use daemon-owned sessions and the shared core loop. The daemon worker owns durable history while its core session manager stays in memory; print has an inline child host, whereas RPC construction does not supply that host. Delegation and persistence therefore depend on the embedding. Provider/model selectors can change, including a live daemon target; the repository revision does not pin exact remote weights or reveal provider adaptation. RT-RTE-daemon-session, RT-CMP-provider-model.
The effect boundary is the executing user's OS permissions. Kernel and worker processes separate lifecycle responsibilities, not security authority. Abort or child disposal does not reverse prior shell/file effects. Recursive depth and sibling-name checks constrain supported child admission, not arbitrary subprocesses or daemon root-session creation. These distinctions matter for any scenario that treats a tool schema, process boundary or recursion limit as containment. RT-CLM-user-permissions, RT-RTE-child-spawn.
Delegation, checking and delivery
Children receive selected tasks in independent working contexts; explicit collection, messages and files carry results back rather than automatically copying working transcripts. Factory adds a symbolic state-machine controller around those model loops. It validates specifications and subagent references, binds outputs, evaluates guards and limits progression. Task text remains model-interpreted. Stored specifications can start later runs, but the live run registry belongs to the Python kernel; losing that registry is not durable workflow replay, and daemon children can survive it. RT-RTE-child-spawn, RT-RTE-factory-run, RT-BAP-machine-guards.
The bundled builder provides one representative connected workflow. When factory is enabled and lookup selects the bundled machine rather than a same-named stored definition, a scope child is asked to inspect the working tree and produce an outline. Writers receive individual outline items; a collector receives their joined sections. The requested five-line outline bound is prompt guidance, while the eight-item expansion cap is executable. Required ports reject missing or unusable upstream output. Completion establishes control-flow quiescence, not accurate counts, sound CI interpretation or report truth. Status exposes only a 200-character answer preview; full child collection and user presentation are separate steps. On a malformed outline the run can pause with an error; resume continues remaining progression but does not redo the failed state. Redo requires a fresh run. WF-RTE-builder-report.
Checking is not uniformly absent or uniformly semantic. Optional autonomous shell gates classify configured command success as status zero without error or timeout, then control continuation, stopping and headless failure reporting. A cached unchanged-workspace failure is not a new measurement. The PR-manager machine instead requests model review of merge-blocking defects and test gaps, routes a false verdict to fixing and a true verdict to monitoring, and returns fix reports to review. JSON capture and strict boolean guards enforce representation and state selection, not reviewer correctness; the template does not automatically merge a PR. EPI-RTE-shell-check, EPI-RTE-gate-disposition, EPI-RTE-pr-review, EPI-RTE-port-capture, EPI-RTE-factory-guard.
What survives and reaches later work
Prime Agent retains several distinct kinds of operating material. Local/global JSON harness stores contain supplemental prompt notes, memory text, skill call contracts, reusable subagent templates and factory specifications. Facts and policies reach models as interpreted context; symbolic machine parts reach code as progression constraints. A harness skill contract points to a callable rather than installing its package. Full stored subagent text can become a later factory child's task. Initial populated-store provenance is not established by a source label or version counter. MEM-OBJ-harness-entry, MEM-OBJ-skill-contract, MEM-OBJ-subagent-entry, MEM-BAP-subagent-template, RT-OBJ-factory-machine.
Automatic digest delivery merges global and local entries and ranks compact hints using active-goal and recent-text terms. Defaults retain three entries per kind with 140-character snippets and ten recent refinement events; these are not a universal token budget. Delivery is a user-role context row, not base-system enforcement. Query terms are excluded from its freshness fingerprint, so task drift alone need not reselect hints. Requested inspection supplies full entries or search results separately, and Python return objects reach the model only as exposed by the cell. Thus a relevant clipped hint is neither full-source delivery nor authenticated knowledge. MEM-RTE-digest-delivery, RT-BAP-harness-digest, MEM-RTE-harness-inspection.
Conversation summaries and executable checkpoints serve different consumers. Compaction replaces a selected prefix with model-written summary and retained tail; optional branch navigation can carry an abandoned-path summary into another selected path. Faithful compression is requested, not semantically guaranteed. Namespace checkpoints serialize eligible Python values for later restoration before bootstrap. A manifest or persistence notice reports names and losses, not values. Oversized or unserializable values can be skipped; post-compaction pruning can delete oversized live names when the kernel is running, a snapshot target exists and artifact replacement succeeds. Process survival therefore does not entail value survival, and restart is best-effort revival rather than exact replay. MEM-OBJ-compaction-summary, MEM-RTE-branch-navigation, MEM-OBJ-kernel-state, MEM-OBJ-kernel-manifest, MEM-RTE-kernel-revival, RT-RTE-context-compact.
Retained adaptation and the improvement claim
The strongest supported improvement-directed contribution is a wired trajectory-to-edit-to-later-consumer path. Manual, queued or automatic refinement supplies a model with bounded session experience, state and history. Software parses proposals, rejects inadmissible edits, checks the planning baseline against reloaded state under lock and applies permissible changes. Local is the default scope; global is explicitly requested. Reserved-base protection applies to this refinement route, and factory create/update is gated again at application while deletion remains available. Direct Python upsert and unrestricted filesystem writes are alternate paths, so these protections are not all-writer invariants. RT-RTE-harness-refine, RT-RTE-python-effects.
Applied entries persist and can reach later foreground or child models. Full rationale and before/after snapshots support audit and rollback, but the next planner and digest receive reduced projections. The event's outcome is copied from the proposal's expected outcome, not measured after application. State mutation and later audit/history writes are not one transaction. These mechanisms support durable trace-fed adaptation, including summary production for later continuation, without establishing improved future capacity or accepted new factual knowledge. Structural admission licenses a store transition; it does not validate truth or predicted benefit. MEM-OBJ-refinement-result, EPI-CLM-harness-improve, profile verification.
The unresolved question is whether evidence-responsive edits actually improve future task effectiveness. No workload, metric or comparison is supplied. Organization mutation and a later guidance channel are wired; actual objective-relevant evidence use by a particular proposal, subsequent behavior causally depending on the edit, and successful improvement remain unobserved. This supports an improvement-directed self-change affordance, not exercised or successful self-improvement. External reports and PR fixes alone are work products, not changes to Prime Agent's own organization. RT-RTE-harness-refine, EPI-RTE-pr-review, EPI-CLM-harness-improve.
Theory-building, reflection and autonomy do not follow as a combined grade. The accepted account establishes editable guidance and requested criticism, not a demonstrated formulated-theory episode satisfying localized content, content-dependent consumption, content-directed criticism and retained criticism shaping iteration. In particular, delivery alone does not establish consumption, and projected history alone does not establish reason-based iteration. Reflection's exercised causal directions and reflective or autonomous theory-building remain unestablished. Separately, computation can schedule optional autonomous continuation and admit refinement without obligatory per-edit approval; that wiring is not proof that every internal theory-building operation is performed autonomously. Improved capacity from criticism of a consumed theory is likewise unestablished, independently of the positive trace-fed update paths. RT-RTE-harness-refine, MEM-OBJ-refinement-result, EPI-RTE-gate-disposition, epistemic conclusion.
For an open coding/research scenario, the inspected architecture supports broad execution, selective context retention and bounded orchestration. For a scenario requiring verified accumulated knowledge, exact workflow replay or demonstrated continual improvement, the available evidence does not establish those stronger guarantees. Search imports retain selected assertions and URLs with unknown upstream warrant; neither retrieval nor persistence supplies acceptance. Assessments would change with linked task traces showing proposal provenance, actual checking, retained edits and later behavioral dependence, plus independently defined future-task comparisons. Faithfulness measurements would change the summary assessment; complete control/recovery traces and durable factory control state would change the replay assessment. No new source reads were performed for this synthesis. EPI-RTE-web-search, RT-RTE-factory-run, report verification.
Limitations
| Limitation | Affected SRC IDs and record citations | Inspected boundary | Conclusion prevented | Evidence that would resolve it |
|---|---|---|---|---|
| No execution or causal comparison | SRC-1; RT-RTE-print-prompt, RT-RTE-harness-refine, EPI-CLM-harness-improve, WF-RTE-builder-report | Frozen Rust/Python source and shipped instructions only | Task success, exercised criticism/acceptance, later behavioral dependence and improved capacity or successful self-improvement | Provenance-linked execution artifacts and objective-defined comparisons with attribution at the intervention's actual scope |
| Retained controls and scheduled admission are incompletely traced; carried report-verification limit | SRC-1; RT-RTE-daemon-session, RT-OBJ-session-context, memory inventory, report limits | Goals, quota parks and heartbeat/cron prompt admission are identified; creation, persistence, cancellation and restart lifecycle are incomplete | Exhaustive memory-route inventory and system-wide recovery, curation or withdrawal comparisons | Records tracing each control's complete writer, retention, admission, cancellation and restart consumers |
| Variant coverage and recovery are incomplete; carried report-verification limit | SRC-1; RT-RTE-skill-install, RT-RTE-release-update, MEM-RTE-branch-navigation, RT-RTE-factory-run, report limits | Selected loader/body expansion, update staging, branch selection and builder/PR-manager paths; incomplete precedence, coordinator recovery, fork copying and other variants | Generalized handoff, replay, withdrawal or delivery guarantees across resources, forks, transports or machines; guaranteed stopping of every child when deletion fails | Variant-specific admission/recovery records, fork artifact ownership and coordinator traces, including failed deletion outcomes |
| Resource admission retains an unresolved cross-report characterization | SRC-1; RT-RTE-skill-install, epistemic ledger, report verification | Runtime calls the whole route afforded; epistemic calls resource consumption implemented; both retain loader/rejection limits | One unqualified implementation status for the whole installation route; neither absence nor observed use follows. Verification dismisses this as a blocker, but the reconciliation's unresolved conflict remains carried | Explicitly separated whole-route and consumption-subfunction records with agreed scope and loader/rejection coverage |
| Retained child-answer classification is incomplete; carried profile-verification limit | SRC-1; RT-RTE-child-spawn, WF-RTE-builder-report, profile limits | Requested later collection is classified, but distinct child-answer units are missing for storage, form, lineage, authority and write agency; axes remain partial | Complete per-object comparison of retained child answers, including persistence substrate and authority at the collecting consumer | Accepted records and separate classification units tracing those answers through retention and later collection |
| Initial contents, checkpoint values and provider state remain unresolved | SRC-1; MEM-OBJ-harness-entry, MEM-OBJ-kernel-state, RT-CMP-provider-model | Empty-store and checkpoint mechanisms inspected; actual populated stores/payloads absent from evidence; remote internals not frozen | Authenticated initial provenance, classification of arbitrary restored values, exact remote model identity or provider-side learning assessment | Registered store/payload artifacts with producer and consumer lineage, and pinned inspectable provider evidence |
| Semantic fidelity, full reason read-back and retrospective withdrawal are not established | SRC-1; MEM-OBJ-compaction-summary, MEM-OBJ-refinement-result, MEM-RTE-digest-delivery, MEM-RTE-harness-inspection | Summary instructions, clipped delivery, projected history and future-load mutation inspected | Guaranteed claim preservation, complete reasons reaching automatic readers, continuous relevance refresh or recall of already-delivered copies | Paired source/summary and full-entry/delivery artifacts, consumer traces, and explicit refresh/withdrawal protocols |