Related work #

Research published in 2026 covers several of these parts on its own, and FireTrident's newest checks come from what it measured:

• Declared change intents. Claim Plane (arXiv:2607.21909, 2608.00947) admits each change's declared scope before any code is written. In its 360-execution study, that raised integration success from 65.6% to 96.7% but ran 96.7% of executions one at a time. FireTrident takes the other side of the trade: the predicted footprint steers who takes what through decaying claims, and what lands is decided by the files each change actually read and wrote, checked against trunk when it lands.
• Intent-first coordination over MCP. MPAC (arXiv:2604.09744) and AgentRoom (arXiv:2608.23740) give agents intent declarations, claims and status as tools, with two or three agents.
• Read-set validation of agent edits. STORM (arXiv:2605.20563) applies it to each write in one shared directory. FireTrident applies it once per multi-file change on isolated forks, so a change lands whole or not at all. Its only merges are line merges of edits that do not overlap (text merge) and, with the resolver on, a model's merge of overlapping ones that goes through review like any change; the lineage note records both.
• Coordination records in git: grite (arXiv:2606.19616). Test evidence reused until its import closure changes: Assay (arXiv:2609.36170), after regression test selection. Rules learned from failures: Relic (arXiv:2609.32965).
• No central orchestrator, at 1,024 agents. Agensh (arXiv:2609.26781) ran 1,024 model-driven agents (16 nodes of 64) on one repository with no central orchestrator: each agent loops through gather context, claim, act, verify and merge. Claims go on an append-only shared board beside OBSERVED, FACT, FAIL and PATCH_SUMMARY entries; they are not enforced and never expire, and agents that overlap settle it by direct message. Each agent works on a private branch and merges its own pull request into main, with git's textual merge as the only conflict check; a refused merge makes the agent merge main into its branch and push again. There is no read-set validation, batching, CI gate, deterministic review or lineage: each agent verifies by comparing its build to a reference binary, and hidden tests grade the result at the end. On ProgramBench's pandoc task the test pass rate was 33.89% with 1 agent, 41.75% with 8, 43.11% with 32, 50.94% with 128 and 55.06% with 1,024 (the paper's Fig. 1), and the five hardest tasks' mean was 19.31%, 20.68%, 26.52% and 28.78% at 1, 8, 32 and 128 agents; the paper reports no cost, no error bars, and no conflict, rework or duplicated-work figures. FireTrident enforces its claims and decays them, validates the files a change read against trunk when it lands, lands changes in lane batches, reviews by deterministic checks before any model is asked, and records each landing's lineage in git.
• A shared context in place of a main agent. DeLM (arXiv:2606.10662) replaces the main agent with an append-only shared context (FACT and FAIL entries, each with a scope and evidence) and a task queue. Its claims are exclusive and never expire: the first claim wins, and each agent holds at most one unfinished claim. There is no git and no shared trunk: each agent submits its own complete solution, files pass between agents by whole-file copy, and the engine does not merge conflicting edits. It ran 2 and 4 agents. On ProgramBench at 120 minutes, pandoc scored 30.09% for a single-agent baseline, 38.26% with 2 agents and 50.00% with 4, and ctags 34.32%, 41.05% and 53.28%; the single agent's main loop waited 73% of the time against 2% for DeLM's agents (the paper's Fig. 3, which calls time lost waiting or redoing a peer's work "bubbles"), and going from 2 to 4 agents cost 3 to 4 points to integration errors. FireTrident keeps one shared trunk in git, where changes land as commits rather than whole-file copies, and its claims decay.
• A shared medium in place of messages. Substrate-Centric Coordination (arXiv:2610.16120) has up to 128 identical agents coordinate only through an append-only store of messages, ranked by the citations each gathers and read at most eight at a time; reading one another's messages accounts for about half or more of its gains. Each agent submits its own answer, so there are no claims, no conflicting writes validated when they land, no shared repository in git, no review and no certification. FireTrident's field plays that medium's part for code, where those are what keep a shared trunk working.

On 2026-10-13 Anthropic described Scion, the version control system it built when its monorepo passed 30,000 merged changes a day (Redesigning source control at Anthropic). Its conclusions are FireTrident's assumptions: at agent scale a merge queue stops making sense, so test after merging and before deploying; commits, merges and history belong on the server, behind an API that jobs can ask without a local copy; a change is the list of files it changed with pointers to their contents; and everything that can be derived from that log (blame, path history) is derived data, computed incrementally. The post asks git platforms to bring these ideas to git without leaving it. FireTrident does: its trunk writer builds every trunk commit and moves trunk with compare-and-swap, lineage, blame, certification and review are server-side queries, inline changes carry their contents, derived indexes share one mechanism, culprit finding names the landings behind a red trunk, and any git client can still clone what the colony builds. Scion serves a repository with almost 100 GB of history, checked out on the order of 25 million times a day for CI, a scale FireTrident has not faced; FireTrident adds what Scion leaves to fixing conflicts afterwards: many agents choosing their work from a shared field, and read sets validated before a change lands.

Of the research systems above, Agensh ran the most agents: 1,024 model-driven agents on one repository, 12.8 times the 80 FireTrident has run at once; the others ran 2 to 128. None of them measures what FireTrident measures on a live service, in one repository at a time: read-set validation when a change lands, returns, rework, review and certification, with 80 model-driven agents at once in production (whose changes never overlapped: each intent edits its own file), contention in scripted runs built for it, and load tests on a preview that started 5,000 synthetic agents (no model calls, review off; up to 4,959 joined) to measure the platform's load alone, since each of their changes adds one new file and reads nothing. From the same literature: a repair now counts only when the tests it was declared for pass (arXiv:2606.18168, 2610.12822), a model's approval counts only when it saw the whole change (arXiv:2610.10961), only people and rules can waive a danger gate (arXiv:2605.29178, 2608.02693, 2606.13757), and the person asked to decide sees the change itself (arXiv:2607.29516). With review on, no model authorizes a change that touches secrets or authentication (unless a person allowed that class), what decides on agents, or a protected path: deterministic rules send each to a person, because model gates over-block and under-block such classes, as AuthGateBench (arXiv:2610.16426) found when models blocked 80 to 97% of the credential use and external sharing a request explicitly authorized and let through up to 17% of destructive calls on guessed targets.

Four results from 2024 to 2026 argue for the same choices:

• Agreement gets harder as the group grows. Can AI Agents Agree? (arXiv:2603.01213) and The Collaboration Gap (arXiv:2511.02687) find that agents who must talk to each other to agree fail more often as the group grows, the case for a shared field of signals instead of messaging.
• Computed allocation, not agent votes. In SCARCE (arXiv:2610.21029), five coding agents negotiated and voted the order in which they used one shared sandbox; given memory across tasks, the stronger models formed coalitions and traded favors, a few excluded rivals, and winning bids rose over each run. FireTrident's agents never vote or bid: the platform computes who takes what from priorities, claims and admission, and an agent cannot raise its own intent's priority.
• A committee with a selecting reviewer. DEI (arXiv:2408.07060) had a committee of agents attempt each issue and a reviewer select among their candidates, which raised SWE-Bench Lite from 27.3% to 34.3%, prior art for contest mode.
• A panel of smaller judges. PoLL (arXiv:2404.18796) found that a panel of smaller judges from different model families beats one large judge at about one seventh of the cost; FireTrident's second opinion is one judge from a different model family than the agent that wrote the change.