Anthropic documents support for mid-conversation system messages and tool changes in Claude, potentially allowing agent harnesses to reconfigure instructions and available tools within an ongoing conversation.
state: watchingheat: lowuncertainty: mediumknownscott: mediumllm-apis agent-harnesses long-running-agentsAnthropic
What is this?
Anthropic’s Claude Platform documentation includes a page titled “Mid-conversation system messages and tool changes,” with a code snippet using a tool-change beta. Amazon Bedrock’s documentation corroborates that mid-conversation system messages can update instructions without resetting prior context, including introducing constraints, refreshing authoritative context, and changing tool guidance. The supplied Claude snippet declares the full tool set up front to preserve a cached prefix, so it does not establish unrestricted addition of new tools mid-conversation; exact tool-change semantics and supported models remain unclear.
Why it matters to Scott
The radar already tracks this development in radar:claude-mid-conversation-system-messages, including the unresolved question of reliable steering without rebuilding context. It bears directly on Scott’s Prompt Interrupt Architecture—delivering new intent into live agent sessions—and Prefix-Caching Economics, providing an API mechanism worth testing, but the supplied evidence does not establish unrestricted tool additions or supported models.
ip:source.prompt-interrupt-architecture-ebookip:concept.prefix-caching-economicsradar:claude-mid-conversation-system-messagesradar:concept.context-engineeringradar:concept.prompt-caching
queries asked of Scott's wikis
- agent harness runtime instruction updates and tool availability
- long-running agents authoritative context refresh
- dynamic tool routing versus fixed tool declarations
- prompt cache stability harness reconfiguration
- system instruction precedence agent control boundaries
Measured heat
now 0 pts/hpeak 3 pts/hcomments 0/hpeers p77momentum: steady2 platformsage 786h
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: p11 vs 519 stories at the 720h mark (now 786h old) — behind addom-local-coding-harness (0.5x)
Evidence (3) — ⭐ canonical anchor
Interpretation history
2026-10-11T11:08:38Z
Anthropic's claude.dev blog post on context engineering for Claude 5 generation models is now linked via HN, but the blog content itself has not been fetched; the open questions about tool-change semantics, supported models, and cache behavior remain unresolved. Engagement remains negligible and the case is over a month old.
2026-10-11T10:35:40Z
evidence attached: hn.story.50041438 — Anthropic's own claude.dev blog publishes context-engineering rules for Claude 5 generation models, directly relevant to mid-conversation reconfiguration.
2026-09-10T23:42:34Z
This look adds no evidence beyond the already routed documentation discovery; live instruction updates remain a relevant implementation lead, not validation of general runtime tool reconfiguration. Cool attention while retaining the open questions about supported models, tool-change semantics, and cache behavior.
2026-09-08T22:45:44Z
The grounded documentation supports mid-conversation instruction updates as a concrete mechanism to test for live-agent control, but does not establish unrestricted tool additions or cache-preserving reconfiguration. This look adds no substantive evidence beyond the previously routed documentation discovery; the HN link and reconstructed documentation title are not independent validation.
2026-09-08T22:37:59Z
grounded: known/medium — The radar already tracks this development in radar:claude-mid-conversation-system-messages, including the unresolved question of reliable steering without rebui
2026-09-08T22:32:29Z
case created — A directly linked first-party API artifact establishes a distinct reconfiguration primitive rather than another echo of the computer-use API release.
Decision trace
- 10-11 22:08repriceAnthropic's claude.dev blog post on context engineering for Claude 5 generation models is now linked via HN, but the blog content itself has not been fetched; the open questions about tool-change
- 10-11 21:41attention_routeNew API capability directly relevant to Scott's agent harness work; can be included in the next briefing for evaluation alongside other tooling updates.
- 10-11 21:35attention_candidateattach
- 10-11 21:35attachAnthropic's own claude.dev blog publishes context-engineering rules for Claude 5 generation models, directly relevant to mid-conversation reconfiguration.
- 10-11 21:35propose_attachAnthropic's own claude.dev blog publishes context-engineering rules for Claude 5 generation models, directly relevant to mid-conversation reconfiguration.
- 09-20 15:23review_dormantscheduled targets exhausted or 28 quiet days
- 09-20 15:23drop_targetsquiet through full ladder or over cap 8
- 09-11 09:42repriceThis look adds no evidence beyond the already routed documentation discovery; live instruction updates remain a relevant implementation lead, not validation of general runtime tool reconfiguration. Co
- 09-11 09:42alert_silentThe documentation discovery was already routed, and the staleness trigger brings no new API detail, implementation result, or actionable constraint. Follow-up can wait for the normal briefing rather t
- 09-11 09:42alert_routeThe documentation discovery was already routed, and the staleness trigger brings no new API detail, implementation result, or actionable constraint. Follow-up can wait for the normal briefing rather t
- 09-09 08:45repriceThe grounded documentation supports mid-conversation instruction updates as a concrete mechanism to test for live-agent control, but does not establish unrestricted tool additions or cache-preserving
- 09-09 08:45alert_silentThe documentation discovery already received an alert route, and this reevaluation contains no new release, access detail, implementation result, or constraint that warrants another interruption. Supp
- 09-09 08:45alert_routeThe documentation discovery already received an alert route, and this reevaluation contains no new release, access detail, implementation result, or constraint that warrants another interruption. Supp
- 09-09 08:42alert_shadowThe linked first-party documentation identifies an API mechanism directly relevant to Scott’s live-agent instruction updates and harness reconfiguration, worth inspecting today before building around
- 09-09 08:42alert_routeThe linked first-party documentation identifies an API mechanism directly relevant to Scott’s live-agent instruction updates and harness reconfiguration, worth inspecting today before building around
- 09-09 08:37groundThe radar already tracks this development in radar:claude-mid-conversation-system-messages, including the unresolved question of reliable steering without rebuilding context. It bears directly on Scot
- 09-09 08:32createA directly linked first-party API artifact establishes a distinct reconfiguration primitive rather than another echo of the computer-use API release.