Research Agenda: Verified Outcomes with Minimal Effective Teams

Objective and Falsifiable Hypothesis

The central research question is whether a small operator can solve previously unsolved problems, or achieve independently verified outcomes using materially fewer resources than a credible reference approach, by coordinating the smallest effective team of autonomous agents.

Falsifiable Hypothesis: Complementary specialization, persistent shared evidence, selective model escalation, and deterministic review gates may improve verified outcomes relative to total compute expenditure. Conversely, uncoordinated scaling introduces duplicate tool calls, correlated model hallucinations, compounding errors, and runaway spend. The research methodology explicitly accommodates the null hypothesis: where a single strong reasoning model or a minimal pair outperforms an adaptive multi-worker fleet.

• Outcome 1: Externally Validated Novelty: A new, verified solution to a difficult problem evaluated against rigorous domain literature and independent human or mechanical checkers.
• Outcome 2: Measured Efficiency Reproduction: Reproduction of a known result achieving parity with fewer total tokens, lower wall-clock duration, or reduced financial expenditure.
• Outcome 3: Documented Failure & Bounds: Negative results where multi-agent scaling added coordination latency or cost without improving accuracy, establishing clear operational boundaries.

The role of echo.cc in research

echo.cc supplies durable messages, task leases, advisory local file reservations, and shared repository-path leases for claims that use the same operator-assigned repository scope. Offline tests cover cross-daemon overlap exclusion, disjoint claims, serial worktree integration, and lease recovery after daemon and broker restarts. These tests do not establish live agent reachability or improved coding-task success. A research harness would need to add experiment bookkeeping, complete cost accounting, independent verification, and retention of unsuccessful attempts. Those research capabilities are proposed designs, not shipped guarantees.

• Independent checks: Scientific validity depends on suitable checkers and domain experts. Agreement between agents is not proof.
• Retain failures: The proposed protocol would record failed branches, discarded approaches, and retries alongside successes.
• Storage boundaries: The local coordination interface uses directory queues and atomic writes. Process recovery and queue durability depend on the host platform and underlying filesystem; experiments must verify their records after interruptions.

Measurement & Resource Accounting Protocol

Rather than collapsing complex multi-agent behavior into a single opaque score, the proposed experiments would measure eight dimensions:

Empirical Status vs. Proposed Experiments

We maintain strict transparency between operational system demonstrations and future scientific research:

• 26-Worker Operational Demonstration (Observed 2026-09-26): A point-in-time process snapshot on host haywire recorded 26 owned live CLI runner processes (5 existing agy workers and 21 dogfood workers instructed to perform read-only audits, running with normal permissions) with 21/21 task acknowledgements and findings under sender rate limits (30 msg/hr burst 10). These are operational observations; it is not a claim of 26 simultaneous inference streams, multi-vendor model diversity, or scientific problem-solving.
• Staged Research Experiments (Proposed & Deferred): Future research will test single-agent baselines against structured small teams on reproducible, held-out benchmark tasks before progressing to prospective open problems. Scaling to dozens or hundreds of workers is an optional future capacity evaluation, not an inherent marker of success.

Measured: 26 API workers and fan-out (2026-09-29)

Measured on haywire on 2026-09-29 between 01:04 and 01:21 UTC, on the live home net, as ordinary members in one campaign. Every worker was ecc peer work calling Gemini gemini-2.5-flash-lite with --max-tokens 128 --timeout 60s and an --expect token. A task counts as verified only when its result-v1 record decodes strictly, hashes to the digest in the task's evidence, and its detail equals the expected token.

Twenty-six API workers ran at once, one task each. One ecc ls --json sample, taken at 01:17:41.351Z, lists all 26 working. The view below is built from that sample by the harness; the 10 other members on the net are reduced to a count. All 26 tasks verified, and no API member was left on the net afterwards. An earlier round also verified 26 of 26, but its best sample showed only 5 of them: ecc ls reads a roster echod rewrites on its own schedule, and a member that lives about 2 seconds can miss every refresh.

# ecc ls --json sample at 2026-09-29T01:17:41.351Z during demo-c26-t26-r2
# 26 api members shown; 10 other members on the net omitted
MEMBER                       STATE    REACH UP   TASK                       VERIFIED RESULT DIGEST
haywire.api.run-0yx5e085     working  turn  3s   01M3NBT6TF0REXHY4VHH8NQS7M yes      01M3NBT8R4FTJMZ9JX0YX5E085 sha256:9ce9c1e19258...
haywire.api.run-24cr7nz6     working  turn  3s   01M3NBT7N9E9BE8WHQ9E7SF6RN yes      01M3NBT8RD0FJD8G5A24CR7NZ6 sha256:766c392a354e...
haywire.api.run-2d0v4w1h     working  turn  3s   01M3NBT7GNG4965BRP9Z8XQXXP yes      01M3NBT8RAFNGQMY7Y2D0V4W1H sha256:66dc9c974fed...
haywire.api.run-2j2dpznh     working  turn  3s   01M3NBT6M25BHTG61CVTPMRRMN yes      01M3NBT8R4TMEXRRSQ2J2DPZNH sha256:bcc1f7fdca1a...
haywire.api.run-302fkygw     working  turn  3s   01M3NBT6H73V1PXDMWH74R7862 yes      01M3NBT8QT2Q39N2P3302FKYGW sha256:2217e6b14fef...
haywire.api.run-3w110zq8     working  turn  3s   01M3NBT7EAKQN6NCBE8W9GF96S yes      01M3NBT8RN1KD01QDH3W110ZQ8 sha256:887bf7b05223...
haywire.api.run-56vfpa8b     working  turn  3s   01M3NBT7QE8410WSM2DGWRSJ9Z yes      01M3NBT8RGWBSCB7VG56VFPA8B sha256:d83fe87f0a0e...
haywire.api.run-5wm31zfa     working  turn  3s   01M3NBT715Z58Z6E8795HWDR9T yes      01M3NBT8RHXJ7ZTNY85WM31ZFA sha256:1d88644a6bad...
haywire.api.run-63b84nrv     working  turn  3s   01M3NBT7CCT76VT6TBFZNXQRVC yes      01M3NBT8RF57T0RNTN63B84NRV sha256:3ebfb2410dda...
haywire.api.run-cdmfcks3     working  turn  3s   01M3NBT7SRNCSQ9PV28FQQSA1E yes      01M3NBT8RJ3JQY0RB0CDMFCKS3 sha256:9505bb635b95...
haywire.api.run-cn9szbwt     working  turn  3s   01M3NBT738HT3AD8SGZA9MXG0Q yes      01M3NBT8RH6RGT760ZCN9SZBWT sha256:a504a9b5ea6e...
haywire.api.run-fjet0w99     working  turn  3s   01M3NBT79V175C6VFC0B2016QH yes      01M3NBT8RKNWW9A9AJFJET0W99 sha256:8ab6606c7cd9...
haywire.api.run-ggfq77ne     working  turn  3s   01M3NBT6QX511FNQ90TRKCPSW8 yes      01M3NBT8R7YN4TSX98GGFQ77NE sha256:31c2e491b780...
haywire.api.run-h076n0zm     working  turn  3s   01M3NBT6E4WJ49CREPAEVAGZCF yes      01M3NBT8R2MAX106M3H076N0ZM sha256:064ec5f25b7b...
haywire.api.run-h6s86jfk     working  turn  3s   01M3NBT7YAH9ZSNTKE9Z04BM2E yes      01M3NBT8RKA6N5HR9JH6S86JFK sha256:014528091760...
haywire.api.run-hda9jrmk     working  turn  3s   01M3NBT80XST3NG1EZXY4QC4VP yes      01M3NBT8RMTSP8GJGRHDA9JRMK sha256:3ffef4f9a78c...
haywire.api.run-hekd1ajw     working  turn  3s   01M3NBT7W5MEPSF9X07E00D3WB yes      01M3NBT8RQP3P5ZHH7HEKD1AJW sha256:28ea6f4fa427...
haywire.api.run-jryxersb     working  turn  3s   01M3NBT75STQF26DS4C6PVNMV1 yes      01M3NBT8R95M0MCSDZJRYXERSB sha256:0335be41ae40...
haywire.api.run-kxjnw92w     working  turn  3s   01M3NBT69DQZPY9B6K9BT9QRYB yes      01M3NBT8RC9Z0SDX9QKXJNW92W sha256:6f5d713a69c2...
haywire.api.run-syck7nxq     working  turn  3s   01M3NBT6WPBYD81Z6F8K285GRB yes      01M3NBT8R101FA5M7XSYCK7NXQ sha256:425301901611...
haywire.api.run-tfr6ngaz     working  turn  3s   01M3NBT839MMBNYS3WCQZY8SW5 yes      01M3NBT8RF2A2JS2W2TFR6NGAZ sha256:b4126ee3e12a...
haywire.api.run-tvexh211     working  turn  3s   01M3NBT77MFRNA53SE9K0H04TJ yes      01M3NBT8RG8PYBF82CTVEXH211 sha256:9a23afc573c8...
haywire.api.run-v208m304     working  turn  3s   01M3NBT6BZ15ZJRJZCP59T48ST yes      01M3NBT8R3QMEFGBDCV208M304 sha256:ac566e157d8a...
haywire.api.run-xe2fjm84     working  turn  3s   01M3NBT6NT45QBE0W66EW66Y1G yes      01M3NBT8R0Y0G60CRMXE2FJM84 sha256:92e621c2b821...
haywire.api.run-ypdphzre     working  turn  3s   01M3NBT6Z401AZF7H66CNRQJFK yes      01M3NBT8RHR1M862G4YPDPHZRE sha256:cad7d4532f44...
haywire.api.run-zc9xmjzj     working  turn  3s   01M3NBT7K2M5K2A4CFXPF5D8S2 yes      01M3NBT8RJQDNNHQXVZC9XMJZJ sha256:a818a3471f63...
# after the run: 0 api members on the net; 26/26 tasks verified (result-v1 decodes, digest matches evidence, detail equals --expect)

Fan-out: each run started N workers at the same instant, one per task. Latencies in milliseconds pool every worker across the runs at a level. Before call is register, claim through echod and setup; after call is persisting the result, the finish through echod and ending the session; provider is the model call.

The model call stays near half a second at every level. What grows is the time echo spends around it, because echod handles claim and finish intents one after another: about 0.2 s per worker at a concurrency of 1, about 1.3 s at 26 and about 7 s at 100. All 384 workers in this sweep finished with no provider errors, timeouts, quota refusals, invalid outputs or retries. When 26 workers raced for one task, exactly one won; the losers exited before any model call. Across every run, including superseded ones, one of 1022 model calls returned a neighbouring task's token, and --expect refused it before it could finish a task. The evidence, scripts and raw data are in release-evidence/bench-20260929/ in the repository.

Measured: isolated recovery qualification (6 October 2026 UTC)

A source-built fixture passed nine actual CLI process checks and eight selected source qualifications. It checks private state persistence, advisory hold expiry, task ownership, stale fences and fail-closed recovery using owned loopback resources.

This is developer-run qualification on scripted boundaries and logical local hosts. It does not measure coding-task acceptance, general autonomous recovery or coordination gains. A fresh upstream coding comparison remains planned; source access is a prerequisite for a reviewer rerun, and no public binary release is implied.

Count and checksum evidence · Recovery fixture and coding-comparison protocol · Dated cross-project methods and limitations

Measured: initial synthetic coding diagnostic (6 October 2026 UTC)

A pinned local model attempted eight original function pairs with two seeds under three conditions: one context, ordinary two-worker messaging and Echo coordination. All 48 attempts are retained; none passed pair acceptance. Each condition accepted zero of 16 pairs. Accounting reconciled 138 requests, 126,116 prompt tokens and 32,975 returned tokens, with complete usage for every attempt.

This negative development diagnostic exposed tool-contract failures. It establishes no coordination benefit, equivalence or representative benchmark performance. External paid spend was zero; hardware and electricity are unmeasured. Cost per accepted pair is undefined.

A separately versioned tool-contract revision completed another 48 attempts on the same exposed fixtures: the single context accepted 8/16 pairs; ordinary two-worker messaging and Echo each accepted 0/16. All 385 requests and 13 merge conflicts are retained. Usage was complete: 409,184 prompt tokens and 23,387 returned tokens. These outcomes remain separate from the first run; changed prompts and validation prevent a fresh-holdout or isolated causal claim. A fresh upstream comparative study remains planned.

Aggregate results and commitments · Protocol and diagnostic scope · Dated cross-project report