CORAL synthesis
Type: agentic-system-analyses/types/agentic-system-synthesis.md
Bounded synthesis
CORAL is an outer runtime for coding-agent experiments. It provisions workspaces, launches external coding harnesses, grades submitted commits, retains experiment and knowledge state, and changes subsequent launches and prompts using results. The model inside each harness chooses implementation steps; CORAL chooses process transitions and configured coordination actions. Its claim to continuously improving agent organizations is supported here as infrastructure and requested knowledge work, not as demonstrated improvement (RT-CLM-agent-organizations).
This account uses inspected implementation and shipped instructions from SRC-1, the CORAL repository at commit 0123dfb939b35228cf2c1fde224cd0561e727408, with applicability cutoff 2026-10-15. The functional whole-system boundary includes shipped callers, templates, evaluation and retained-state consumers; it does not mean every adapter, grader or frontend component was inspected. External harness/provider internals and deployment behavior are not frozen. No target execution, supplied execution trace or causal experiment establishes activation or benefit. The boundary and Source register define these limits. This synthesis adds no source inspection.
From objective to candidate
An operator supplies a task, seed code and packaged grader. The ordinary coral start -c task.yaml path structurally validates the configuration, then a manager provisions project state, agent worktrees, shared-state links and external harness processes. Structural validation is not baseline grading. Agent identity ties together a branch, role and attempt metadata; island identity selects the shared knowledge neighborhood. Instructions and runtime settings reach the harness through files and launch arguments. The harness owns inner model/tool turns, including its own context handling (RT-RTE-agent-launch, RT-OBJ-task-config).
Agents work on separate checked-out code while sharing notes, skills, roles and attempt records through files. Generated guidance prescribes plan → edit → eval → repeat and directs agents to existing knowledge. Specialist work is afforded through seeded definitions, but spawning, context separation and return transport belong to the external harness. A specialist can change shared files as well as return a summary, so the effects are not confined to a return value (RT-BAP-generated-instructions, RT-RTE-specialist-delegation).
Independent worktrees are not an unconditional security boundary. Codex bypasses native approval and sandbox checks on inspected fresh/resume paths; stronger containment depends on an optional outer wrapper and an enforcing deployment. Mounts and custom imports can change runtime machinery. At this commit, the advertised Pi manager path has a static interface mismatch: the manager passes sandbox, but the adapter's explicit start signature does not accept it, even when its value is None. This predicts failure before subprocess launch, not an observed failed run or a defect in the external Pi harness (RT-CLM-isolation-scope, RT-RTE-machinery-admission, RT-CLM-pi-launch).
Submission, measurement and delivered feedback
Calling coral eval admits current edits through a producer-side queue check, commits code, checkpoints shared state and writes a pending attempt. Neither a passing score nor human approval gates that commit. The checkpoint records existing shared bytes before grading; it does not certify their claims and can fail without stopping submission. The daemon grades a detached checkout of the selected commit, removes it and finalizes score, status and feedback. Snapshot stability separates evaluation from concurrent live edits, not from mutable external dependencies or hostile effects (RT-RTE-eval-submission, MEM-OBJ-checkpoint, RT-RTE-grader-execution).
The representative workflow is circle packing: a program returns 26 centers, radii and a claimed sum for a unit square. The inspected grader checks array shapes, NaNs, negative radii, boundaries and overlap, with numerical tolerances, then recomputes the sum and divides it by the reference value 2.635977. It does not simply trust the candidate's sum. This supports a bounded task measurement, not global optimality or the truth of an explanation for a score. The exact shipped example selects docker-opencode, unlike the generic default-Claude launch trace; no end-to-end execution of that configured example is established (workflow account, RT-RTE-grader-execution).
The information reaching the author is narrower than the check's internal result. An overlap can produce score zero and diagnostic details, but the wrapper omits those details from its normal explanation. A log pointer does not establish that the missing reason was logged or read. A CLI wait timeout instead means still pending, with a route to wait again; it is not grading rejection. Locally delivered output is a committed candidate and inspectable terminal attempt, which may record failure. Optional export creates a caller-selected Git branch without checking score/status; it is neither automatic winner promotion nor deployment (workflow annotations, WF-RTE-candidate-export).
Results change continuation and shared knowledge through different routes
Finalized scores have executable consequences. The manager applies stopping and heartbeat rules and can interrupt/resume the committing agent with selected result fields and action prompts. Selection combines broad eligibility/cadence with recipient identity. It does not embed the full knowledge store or guarantee an interrupt for every finalized attempt: batched arrivals can be marked seen while one latest result supplies feedback. Tune budgets are excluded from real-score plateau history, but not universally from every interval counter. Numeric status is also narrow: the first numeric attempt can be labeled improved without a baseline comparison (RT-RTE-heartbeat-feedback, EPI-OBJ-attempt-score).
Shared knowledge takes a separate, model-mediated path. Results, logs, transcripts or research can feed requested explanations, revised roles and reusable skills; retained output can guide later planning or procedures. Markdown bodies supply natural-language guidance, parsed metadata supplies fixed selection rules, and invoked skill scripts supply executable behavior. Files and a shared-state Git history retain these different parts. Merely checkpointing a result is not the same as deriving a lesson (MEM-RTE-knowledge-revision, MEM-OBJ-skills, MEM-OBJ-roles, MEM-BAP-retained-guidance).
Later knowledge access is principally requested: file reads, CLI queries and API calls return listings, bodies or projections. Literal search, date/status filters and model-selected paths are not unsolicited semantic recall. Addressing varies by caller: a nonnumeric CLI note read concatenates bodies; numeric reads omit the requested status predicate; skill detail selection differs between CLI and API. Archiving notes changes generated-index visibility without uniformly removing them from hub enumeration. Supersession metadata and graph edges likewise do not enforce universal withdrawal (MEM-RTE-knowledge-query, MEM-RTE-note-maintenance, EPI-RTE-note-selection).
Research instructions add substantive, but unobserved, review. They request saved sources, per-claim textual-support judgments and an author decision to revise, downgrade or withdraw reliance. Source warrant remains unknown, and no concrete source/output pair establishes whether generated synthesis preserves, derives or conjectures content. The implemented grounding/lint scripts check structure: review-file existence and a supplied verified flag are not a passed semantic review or an experiment recheck. Default advisory exits and optional strict failure do not prevent direct retention or reading. No accepted novel claim or subsequent integration is established by the supplied evidence (EPI-RTE-research-writing, EPI-RTE-synthesis-review, EPI-RTE-confidence-disposition, EPI-RTE-grounding-check, EPI-RTE-note-lint). Requested dedup adjudication and paired with-skill/without-skill tests are additional evaluator alternatives, not an exhaustively assessed acceptance architecture (report verification limits).
Recovery, coordination and the improvement boundary
Steering can rewind descendant worktrees and archive discarded attempts while retaining direct recovery possibilities. Migration uses scored history to relocate an eligible agent, move identity/cadence/attempt artifacts and repoint its shared links; source notes and skills remain behind. Saved conversation can retain source-island information. Neither operation atomically restores or erases code, knowledge, session narrative and external effects (RT-RTE-resume-steering, RT-RTE-island-migration, MEM-OBJ-session-state).
The strongest supported learning contribution is a set of retained update routes, not a measured gain. Result/transcript-to-note/skill/role revision is afforded and has prescribed later readers. Migration additionally implements an automatic arrival summary derived from relocation and recent score state, available to later knowledge consumers. That is a wired trace-fed coordination artifact, not a distilled successful technique or proof of later model dependence. These positives remain valid while actual improved capacity for future action is unestablished; the memory profile's trace-learning classification does not assert such improvement (MEM-RTE-knowledge-revision, MEM-OBJ-migration-arrival, profile verification).
The independent property conclusions are:
- Theory-builder conditions 1–4: Each has operation status uninspected for the requested theory-like knowledge routes. No produced whole theory, content-dependent consumption, attributable content-directed criticism or kept criticism shaping another round was supplied. Editable claim/evidence text affords addressability, but scores selecting candidates do not establish theory criticism (MEM-RTE-knowledge-revision, EPI-RTE-research-writing).
- Learning from criticism: No supplied episode shows criticism of a consumed theory improving future capacity. This withholds a benefit/attribution conclusion, not the afforded revision path (MEM-RTE-knowledge-revision).
- Reflection: Role self-representation and result analysis are requested. An exercised loop connecting changes in the represented aspect to representation updates and representation-mediated later behavior remains unestablished (MEM-OBJ-roles, MEM-RTE-knowledge-revision).
- Autonomy: Process scheduling, grading and migration are wired computational operations without per-action human approval on those paths. Autonomous completion of requested knowledge work or all internal theory-building roles is unobserved; the initial operator request does not demonstrate it (RT-RTE-heartbeat-feedback, RT-RTE-island-migration, MEM-RTE-knowledge-revision).
- Self-improvement: Within CORAL plus its retained guidance over later rounds/sessions of one task, revised skills and roles could change behavior-determining organization toward the independently specified task objective. The supported route is afforded, not exercised self-improvement. No observed chain connects objective-bearing evidence to a particular retained update and subsequent operation causally depending on it; successful improvement needs separate outcome evidence. Better external candidate code alone would not establish CORAL self-improvement (MEM-RTE-knowledge-revision, RT-CLM-agent-organizations).
For metric-directed experimentation, the source supports committed-candidate measurement, scheduling and recoverable handoff. For evidence-sensitive research reuse, it supports durable artifacts and requested review while leaving semantic acceptance, actual reliance and benefit unresolved. An attributed execution linking input evidence, authored update, review/disposition and later consumer would change the exercised-behavior assessment. Controlled later-capacity comparisons would address benefit; independently varying knowledge sharing or a particular update would be needed for attribution to that component. Repairing the Pi interface, preserving rejection details in feedback, and enforcing verdict-aware admission or consistent withdrawal would alter the corresponding architectural limits, not retroactively demonstrate successful operation.
Limitations
| Limitation | Affected SRC IDs and record citations | Inspected boundary | Conclusion prevented | Evidence that would resolve it |
|---|---|---|---|---|
| No execution or causal comparison supplied | SRC-1; RT-RTE-agent-launch, MEM-RTE-knowledge-revision, EPI-RTE-research-writing | Frozen implementation and instructions, not an executed experiment | Actual compliance, exercised theory-building, improved future capacity and causal benefit from sharing | Attributed execution/provenance chain and appropriately controlled later-capacity comparisons |
| Evaluator inventory is partial; carried report-verification limit | SRC-1; MEM-RTE-note-maintenance, MEM-OBJ-skills, MEM-RTE-knowledge-revision, epistemic coverage, declared limit | Retained-knowledge maintenance and skill optimization; requested dedup judge and with/without-skill tests identified | Scores and synthesis review exhaust all evaluators; all maintenance/skill acceptance criteria assessed; bundled autonomous optimizer exists | Function-level assessment of these alternatives, resolution of missing helper dependencies, and execution evidence for activation |
| Host context and provider state are opaque | SRC-1; MEM-OBJ-session-state, RT-RTE-specialist-delegation, RT-CMP-model-endpoints | CORAL pointers, adapters and requested summaries, not external internals | Complete memory inventory, fixed model weights, faithful replay, enforced context budgets or actual specialist isolation | Frozen host/provider implementation or attributed payload/context evidence under specified versions |
| Exact example and adapter parity are not established | SRC-1; workflow account, RT-CLM-pi-launch, RT-RTE-agent-launch | Default launch trace, separate circle-packing Docker/OpenCode configuration and static Pi mismatch | Successful packaged end-to-end run or equivalent behavior across adapters | Corrected Pi interface and inspected/attributed launches for the exact configurations |
| Containment and external effects depend on deployment | SRC-1; RT-CLM-isolation-scope, RT-RTE-grader-execution | Optional wrapping, launch settings, detached checkout and worker timeout | Universal host/private-state protection, complete model-call mediation or descendant cleanup | Specified enforcing OS/runtime envelope and containment/cleanup evidence for covered and alternate paths |
| Measurement and delivered diagnosis have narrower scope than task success | SRC-1; RT-RTE-grader-execution, workflow annotations, EPI-OBJ-attempt-score | Circle-packing checks, recomputed score, status assignment and wrapper feedback | Global optimality, all-task correctness, baseline gain from the improved label or automatic delivery of violated constraint | Domain-specific verifier assurance, benchmark provenance, explicit baseline comparisons and complete delivered diagnostics |
| Shared retention and structural checks do not establish semantic acceptance | SRC-1; MEM-OBJ-checkpoint, EPI-RTE-grounding-check, EPI-RTE-note-lint, EPI-RTE-confidence-disposition | Checkpointing, optional local checks and requested author disposition | True checkpointed notes, authentic/lossless source import, passed semantic review or accepted novel knowledge | Actual paired sources/claims, substantive review and evidence-consuming acceptance decisions; verdict-aware controls for an enforced guarantee |
| Read-back and withdrawal differ across consumers | SRC-1; MEM-RTE-knowledge-query, MEM-RTE-note-maintenance, EPI-RTE-note-selection | Inspected CLI/hub/API, index/move helpers; frontend and some helper internals uninspected | Uniform addressing, universal retirement, lossless structured provenance or complete maintenance coverage | Aligned consumer predicates, structured metadata-preservation evidence and assessment of omitted consumers/helpers |
| Feedback delivery and recovery are not complete transactions | SRC-1; RT-RTE-heartbeat-feedback, RT-RTE-resume-steering, RT-RTE-island-migration | Latest-result selection, ancestry rewind, attempt archival and shared-link relocation | One interrupt per attempt, universal tune exclusion, atomic rollback or source-context erasure | Per-attempt delivery semantics and traces; coordinated restoration/erasure contracts covering code, knowledge, sessions and external effects |