KoboldCpp maintainer concedo claims the newly bundled single-checkbox agent harness (nine tools, a compact built-in prompt) makes basic agentic coding practical on local models without external harnesses; uptake and real-task results from LocalLLaMA users will show whether bundled lightweight harnesses suffice for everyday tasks.
state: watchingheat: highuncertainty: mediumconvergesscott: highagent-harnesses local-inference coding-agentsconcedo
Surfaced 2026-09-27T10:43:47Z — Release notes for koboldcpp-1.122 (published 2026-09-26T09:05:10Z by LostRuins): "NEW: Added an integrated KoboldCpp Agent - KoboldCpp now s — Second consecutive quiet look: +15 pts/+1 comment over 16.5h, entirely inside concedo's announcement thread; the velocity-spike and magnitude-valve triggers are again artifacts (absolute score 151 vs p90 135; 'several platforms' = the actor's own Reddit thread plus his own release notes — one actor, one source line), and measured_heat's 'accelerating' momentum contradicts the raw ~0.9 pts/h, ~0.06 comments/h long-tail decay. Meaning unchanged — confirmed one-actor release, still zero third-party trial results — so attention holds at low; expiry is the expected next step on a third quiet look, while a first credible trial report or uptake receipt remains the sole material reheat trigger.
What is this?
Per the case, KoboldCpp maintainer concedo has bundled an agent harness into the runtime — enabled by a single checkbox, exposing nine tools behind a compact built-in prompt — claiming this makes basic agentic coding practical on local models without an external harness; the anchor post also carries a plea for community help. Caution: none of the retrieved snippets cover this release directly; they only establish the surrounding 2026 discourse the release lands in — harnesses framed as the decisive pre-wired runtime layer around a model (harness vs assemble-it-yourself framework), a controlled LocalLLaMA experiment showing four harnesses on identical Qwen3.6 27B weights produce measurably different coding output, and an active genre of community-built local coding harnesses (Prime Agent). So the claim fits the landscape but is unverified by these snippets, and no uptake or real-task results are evidenced yet.
Why it matters to Scott
Converges with his workshop-ebook thesis that harness persistence and tool surface are multiplicative capability factors — a first-party local runtime (peer to his Ollama/gamepc stack) now bets that a nine-tool, compact-prompt checkbox harness suffices for agentic coding, the exact lane of his own `ask` --ollama path and directly testable against it with his trace-backed comparison fixture. It also bears on the scaffolding hypothesis: if a fixed universal bundled harness proves sufficient for everyday tasks, bespoke team-owned scaffolding loses part of its claim — and concedo's pending uptake results are dated receipts either way.
ip:source.give-the-agent-a-workshop-ebookip:concept.scaffolding-hypothesisdev:project.askdev:technology.ollamaip:concept.attention-budgetradar:concept.agent-harnessesradar:llama-cpp-rootless-tool-sandboxesradar:marmel-090-local-codingradar:qwen38-27b-local-api-substitutionradar:ship-harness-benchradar:concept.open-source-maintenance
queries asked of Scott's wikis
- bundled first-party agent harness inside local inference server
- minimal harness tool count compact system prompt for small local models
- harness scaffolding variance same model different harness results
- local open-weight models agentic coding viability everyday tasks
- agent loop tool registry while-loop harness architecture patterns
- single maintainer plea for help open source project sustainability
Measured heat
now 0 pts/hpeak 0 pts/hcomments 0/hpeers p0momentum: steady2 platformsage 386h
points/hour across evidence · reading as of 2026-10-12 02:59:37.977291+11:00 · deterministic, not a model opinion
How the heat travelled
pace: p75 vs 1032 stories at the 336h mark (now 386h old) — ahead of ftc-agent-developer-liability (1.0x), behind qwen06b-note8-browser-control (1.0x)
Evidence (2) — ⭐ canonical anchor
Interpretation history
2026-09-27T10:39:19Z
magnitude valve eligible (multi-platform, top-decile engagement) and never alerted; deterministic escalation to deliver
2026-09-27T03:54:39Z
Sixteen more hours of decay inside the same single thread (+15 pts/+5 comments, 2.5 pts/h vs a 19 pt/h peak); the velocity-spike trigger and the magnitude-valve spread reading are both artifacts of that one announcement thread crossing an absolute p90 (136 vs 135), not expanding periphery — still 2 platforms, 1 actor, 0 derivatives, 0 independent implementations. Meaning is unchanged — confirmed one-actor release still awaiting its first credible third-party trial report — so attention cools to low; a second consecutive quiet look with no results trends toward expiry, while any trial report or uptake receipt remains the material trigger to reheat.
2026-09-26T20:51:20Z
Engagement roughly doubled (121 pts/25 comments at ~31h) but every increment sits inside concedo's own announcement thread; new comments are praise and prompt-size commentary (2k tokens vs 10x competitors), not results. The 9.4× velocity spike is single-source amplification of a one-actor story, so no corroboration and no change in meaning: still a confirmed release awaiting its first third-party trial report. Heat stays medium despite the 91.7th percentile rate because momentum is cooling and the periphery is not expanding (two platforms, one actor, no derivatives).
2026-09-26T13:41:36Z
The velocity spike (7× peer baseline, 90th percentile) is contained within concedo's own announcement thread — both evidence objects still trace to a single actor, so no independent corroboration exists yet. At ~23h the comments are reactions (thanks, KV-cache praise, video snark), not results: nobody has reported running the bundled harness on a real task, so the case's deciding evidence hasn't arrived; the move to watching reflects a confirmed live release with an active thread, not new substance.
2026-09-26T09:53:32Z
origin walked (opencode/cheap-glm, conf 0.9): anchor reddit.post.1wqlyp8 -> echo.github.c51a8d0909 by concedo (LostRuins) / KoboldCpp project
2026-09-26T09:33:16Z
grounded: converges/high — Converges with his workshop-ebook thesis that harness persistence and tool surface are multiplicative capability factors — a first-party local runtime (peer to
2026-09-26T09:25:56Z
case created — First-party maintainer release of a usable local agent harness — a concrete artifact on two hot topics with no existing open case.
Decision trace
- 10-03 18:47review_dormantscheduled targets exhausted or 28 quiet days
- 10-03 18:47drop_targetsquiet through full ladder or over cap 8
- 09-27 20:43pushRelease notes for koboldcpp-1.122 (published 2026-09-26T09:05:10Z by LostRuins): "NEW: Added an integrated KoboldCpp Agent - KoboldCpp now s — Second consecutive quiet look: +15 pts/+1 comment ov
- 09-27 20:39repriceSecond consecutive quiet look: +15 pts/+1 comment over 16.5h, entirely inside concedo's announcement thread; the velocity-spike and magnitude-valve triggers are again artifacts (absolute score 15
- 09-27 20:39alert_heldRelease notes for koboldcpp-1.122 (published 2026-09-26T09:05:10Z by LostRuins): "NEW: Added an integrated KoboldCpp Agent - KoboldCpp now s — Second consecutive quiet look: +15 pts/+1 comment ov
- 09-27 20:39alert_routeRelease notes for koboldcpp-1.122 (published 2026-09-26T09:05:10Z by LostRuins): "NEW: Added an integrated KoboldCpp Agent - KoboldCpp now s — Second consecutive quiet look: +15 pts/+1 comment ov
- 09-27 20:20sensor_dirtyvelocity_spike
- 09-27 13:54repriceSixteen more hours of decay inside the same single thread (+15 pts/+5 comments, 2.5 pts/h vs a 19 pt/h peak); the velocity-spike trigger and the magnitude-valve spread reading are both artifacts of th
- 09-27 12:20sensor_dirtyvelocity_spike
- 09-27 08:21sensor_dirtycomment_update
- 09-27 06:51repriceEngagement roughly doubled (121 pts/25 comments at ~31h) but every increment sits inside concedo's own announcement thread; new comments are praise and prompt-size commentary (2k tokens vs 10x co
- 09-27 04:21sensor_dirtycomment_update
- 09-27 03:22sensor_dirtyvelocity_spike
- 09-27 00:21sensor_dirtycomment_update
- 09-26 23:41repriceThe velocity spike (7× peer baseline, 90th percentile) is contained within concedo's own announcement thread — both evidence objects still trace to a single actor, so no independent corroboration
- 09-26 21:20sensor_dirtyvelocity_spike
- 09-26 20:20sensor_dirtycomment_update
- 09-26 19:53promote_anchororigin walk conf 0.9
- 09-26 19:33groundConverges with his workshop-ebook thesis that harness persistence and tool surface are multiplicative capability factors — a first-party local runtime (peer to his Ollama/gamepc stack) now bets that a
- 09-26 19:25createFirst-party maintainer release of a usable local agent harness — a concrete artifact on two hot topics with no existing open case.